[ { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0000.png", "metadata/rollout_000/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0001.png", "metadata/rollout_000/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0002.png", "metadata/rollout_000/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0003.png", "metadata/rollout_000/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0004.png", "metadata/rollout_000/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0005.png", "metadata/rollout_000/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0006.png", "metadata/rollout_000/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0007.png", "metadata/rollout_000/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0008.png", "metadata/rollout_000/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0009.png", "metadata/rollout_000/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0010.png", "metadata/rollout_000/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0011.png", "metadata/rollout_000/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0012.png", "metadata/rollout_000/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0013.png", "metadata/rollout_000/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_000/agentview/0014.png", "metadata/rollout_000/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_000/agentview/0015.png", "metadata/rollout_000/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_000/agentview/0016.png", "metadata/rollout_000/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0017.png", "metadata/rollout_000/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0018.png", "metadata/rollout_000/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0019.png", "metadata/rollout_000/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0020.png", "metadata/rollout_000/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0021.png", "metadata/rollout_000/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_000/agentview/0022.png", "metadata/rollout_000/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_000/agentview/0023.png", "metadata/rollout_000/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_000/agentview/0024.png", "metadata/rollout_000/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0025.png", "metadata/rollout_000/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_000/agentview/0026.png", "metadata/rollout_000/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0027.png", "metadata/rollout_000/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_LEFT, MV_BACK, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0028.png", "metadata/rollout_000/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0029.png", "metadata/rollout_000/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0030.png", "metadata/rollout_000/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0031.png", "metadata/rollout_000/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0032.png", "metadata/rollout_000/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0033.png", "metadata/rollout_000/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0034.png", "metadata/rollout_000/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0035.png", "metadata/rollout_000/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0036.png", "metadata/rollout_000/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0037.png", "metadata/rollout_000/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0038.png", "metadata/rollout_000/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0039.png", "metadata/rollout_000/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0040.png", "metadata/rollout_000/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0041.png", "metadata/rollout_000/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0042.png", "metadata/rollout_000/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0043.png", "metadata/rollout_000/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0044.png", "metadata/rollout_000/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0045.png", "metadata/rollout_000/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0046.png", "metadata/rollout_000/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0047.png", "metadata/rollout_000/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0048.png", "metadata/rollout_000/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0049.png", "metadata/rollout_000/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0050.png", "metadata/rollout_000/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0051.png", "metadata/rollout_000/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0052.png", "metadata/rollout_000/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0053.png", "metadata/rollout_000/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0054.png", "metadata/rollout_000/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0055.png", "metadata/rollout_000/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_000/agentview/0056.png", "metadata/rollout_000/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_000/agentview/0057.png", "metadata/rollout_000/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_000/agentview/0058.png", "metadata/rollout_000/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0059.png", "metadata/rollout_000/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_BACK, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_000/agentview/0060.png", "metadata/rollout_000/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_BACK, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0061.png", "metadata/rollout_000/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_000/agentview/0062.png", "metadata/rollout_000/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_000/agentview/0062.png", "metadata/rollout_000/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0000.png", "metadata/rollout_001/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0001.png", "metadata/rollout_001/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0002.png", "metadata/rollout_001/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0003.png", "metadata/rollout_001/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0004.png", "metadata/rollout_001/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0005.png", "metadata/rollout_001/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0006.png", "metadata/rollout_001/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0007.png", "metadata/rollout_001/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0008.png", "metadata/rollout_001/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0009.png", "metadata/rollout_001/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_001/agentview/0010.png", "metadata/rollout_001/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0011.png", "metadata/rollout_001/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0012.png", "metadata/rollout_001/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0013.png", "metadata/rollout_001/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0014.png", "metadata/rollout_001/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0015.png", "metadata/rollout_001/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0016.png", "metadata/rollout_001/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0017.png", "metadata/rollout_001/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0018.png", "metadata/rollout_001/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0019.png", "metadata/rollout_001/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_001/agentview/0020.png", "metadata/rollout_001/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_001/agentview/0021.png", "metadata/rollout_001/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_001/agentview/0022.png", "metadata/rollout_001/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0023.png", "metadata/rollout_001/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0024.png", "metadata/rollout_001/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0025.png", "metadata/rollout_001/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0026.png", "metadata/rollout_001/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0027.png", "metadata/rollout_001/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_001/agentview/0028.png", "metadata/rollout_001/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_001/agentview/0029.png", "metadata/rollout_001/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_001/agentview/0030.png", "metadata/rollout_001/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_001/agentview/0031.png", "metadata/rollout_001/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_001/agentview/0032.png", "metadata/rollout_001/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_001/agentview/0033.png", "metadata/rollout_001/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_001/agentview/0034.png", "metadata/rollout_001/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_001/agentview/0035.png", "metadata/rollout_001/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0036.png", "metadata/rollout_001/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0037.png", "metadata/rollout_001/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0038.png", "metadata/rollout_001/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0039.png", "metadata/rollout_001/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0040.png", "metadata/rollout_001/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0041.png", "metadata/rollout_001/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0042.png", "metadata/rollout_001/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0043.png", "metadata/rollout_001/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0044.png", "metadata/rollout_001/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0045.png", "metadata/rollout_001/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0046.png", "metadata/rollout_001/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0047.png", "metadata/rollout_001/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_001/agentview/0048.png", "metadata/rollout_001/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0049.png", "metadata/rollout_001/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0050.png", "metadata/rollout_001/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_001/agentview/0051.png", "metadata/rollout_001/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_001/agentview/0052.png", "metadata/rollout_001/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_001/agentview/0053.png", "metadata/rollout_001/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_001/agentview/0054.png", "metadata/rollout_001/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_001/agentview/0054.png", "metadata/rollout_001/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0000.png", "metadata/rollout_002/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0001.png", "metadata/rollout_002/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0002.png", "metadata/rollout_002/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0003.png", "metadata/rollout_002/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0004.png", "metadata/rollout_002/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0005.png", "metadata/rollout_002/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0006.png", "metadata/rollout_002/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0007.png", "metadata/rollout_002/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0008.png", "metadata/rollout_002/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_002/agentview/0009.png", "metadata/rollout_002/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_002/agentview/0010.png", "metadata/rollout_002/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_002/agentview/0011.png", "metadata/rollout_002/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_002/agentview/0012.png", "metadata/rollout_002/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_002/agentview/0013.png", "metadata/rollout_002/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_002/agentview/0014.png", "metadata/rollout_002/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0015.png", "metadata/rollout_002/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0016.png", "metadata/rollout_002/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0017.png", "metadata/rollout_002/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0018.png", "metadata/rollout_002/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0019.png", "metadata/rollout_002/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0020.png", "metadata/rollout_002/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_002/agentview/0021.png", "metadata/rollout_002/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0022.png", "metadata/rollout_002/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0023.png", "metadata/rollout_002/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0024.png", "metadata/rollout_002/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0025.png", "metadata/rollout_002/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0026.png", "metadata/rollout_002/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0027.png", "metadata/rollout_002/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0028.png", "metadata/rollout_002/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0029.png", "metadata/rollout_002/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0030.png", "metadata/rollout_002/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0031.png", "metadata/rollout_002/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0032.png", "metadata/rollout_002/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0033.png", "metadata/rollout_002/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0034.png", "metadata/rollout_002/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0035.png", "metadata/rollout_002/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0036.png", "metadata/rollout_002/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0037.png", "metadata/rollout_002/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0038.png", "metadata/rollout_002/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_002/agentview/0039.png", "metadata/rollout_002/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_002/agentview/0040.png", "metadata/rollout_002/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_002/agentview/0041.png", "metadata/rollout_002/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_002/agentview/0042.png", "metadata/rollout_002/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0043.png", "metadata/rollout_002/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0044.png", "metadata/rollout_002/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0045.png", "metadata/rollout_002/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_002/agentview/0046.png", "metadata/rollout_002/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_002/agentview/0047.png", "metadata/rollout_002/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0048.png", "metadata/rollout_002/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_002/agentview/0049.png", "metadata/rollout_002/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_002/agentview/0049.png", "metadata/rollout_002/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0000.png", "metadata/rollout_003/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0001.png", "metadata/rollout_003/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0002.png", "metadata/rollout_003/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0003.png", "metadata/rollout_003/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0004.png", "metadata/rollout_003/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0005.png", "metadata/rollout_003/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0006.png", "metadata/rollout_003/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0007.png", "metadata/rollout_003/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0008.png", "metadata/rollout_003/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_003/agentview/0009.png", "metadata/rollout_003/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_003/agentview/0010.png", "metadata/rollout_003/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_003/agentview/0011.png", "metadata/rollout_003/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_003/agentview/0012.png", "metadata/rollout_003/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_003/agentview/0013.png", "metadata/rollout_003/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_003/agentview/0014.png", "metadata/rollout_003/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_003/agentview/0015.png", "metadata/rollout_003/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_003/agentview/0016.png", "metadata/rollout_003/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0017.png", "metadata/rollout_003/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0018.png", "metadata/rollout_003/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0019.png", "metadata/rollout_003/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0020.png", "metadata/rollout_003/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0021.png", "metadata/rollout_003/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0022.png", "metadata/rollout_003/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0023.png", "metadata/rollout_003/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0024.png", "metadata/rollout_003/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0025.png", "metadata/rollout_003/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0026.png", "metadata/rollout_003/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_003/agentview/0027.png", "metadata/rollout_003/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0028.png", "metadata/rollout_003/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0029.png", "metadata/rollout_003/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0030.png", "metadata/rollout_003/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0031.png", "metadata/rollout_003/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0032.png", "metadata/rollout_003/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0033.png", "metadata/rollout_003/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0034.png", "metadata/rollout_003/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0035.png", "metadata/rollout_003/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_003/agentview/0036.png", "metadata/rollout_003/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_003/agentview/0037.png", "metadata/rollout_003/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_003/agentview/0038.png", "metadata/rollout_003/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0039.png", "metadata/rollout_003/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0040.png", "metadata/rollout_003/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0041.png", "metadata/rollout_003/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0042.png", "metadata/rollout_003/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0043.png", "metadata/rollout_003/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0044.png", "metadata/rollout_003/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_003/agentview/0045.png", "metadata/rollout_003/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0046.png", "metadata/rollout_003/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0047.png", "metadata/rollout_003/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0048.png", "metadata/rollout_003/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_003/agentview/0049.png", "metadata/rollout_003/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_003/agentview/0050.png", "metadata/rollout_003/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0051.png", "metadata/rollout_003/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_003/agentview/0052.png", "metadata/rollout_003/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_003/agentview/0052.png", "metadata/rollout_003/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0000.png", "metadata/rollout_004/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0001.png", "metadata/rollout_004/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0002.png", "metadata/rollout_004/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0003.png", "metadata/rollout_004/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0004.png", "metadata/rollout_004/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0005.png", "metadata/rollout_004/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0006.png", "metadata/rollout_004/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0007.png", "metadata/rollout_004/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0008.png", "metadata/rollout_004/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0009.png", "metadata/rollout_004/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0010.png", "metadata/rollout_004/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0011.png", "metadata/rollout_004/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0012.png", "metadata/rollout_004/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0013.png", "metadata/rollout_004/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0014.png", "metadata/rollout_004/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0015.png", "metadata/rollout_004/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0016.png", "metadata/rollout_004/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0017.png", "metadata/rollout_004/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_004/agentview/0018.png", "metadata/rollout_004/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_004/agentview/0019.png", "metadata/rollout_004/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0020.png", "metadata/rollout_004/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0021.png", "metadata/rollout_004/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0022.png", "metadata/rollout_004/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0023.png", "metadata/rollout_004/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0024.png", "metadata/rollout_004/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0025.png", "metadata/rollout_004/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_004/agentview/0026.png", "metadata/rollout_004/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0027.png", "metadata/rollout_004/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0028.png", "metadata/rollout_004/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0029.png", "metadata/rollout_004/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0030.png", "metadata/rollout_004/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0031.png", "metadata/rollout_004/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0032.png", "metadata/rollout_004/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0033.png", "metadata/rollout_004/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0034.png", "metadata/rollout_004/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_004/agentview/0035.png", "metadata/rollout_004/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_004/agentview/0036.png", "metadata/rollout_004/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_004/agentview/0037.png", "metadata/rollout_004/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_004/agentview/0038.png", "metadata/rollout_004/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_004/agentview/0039.png", "metadata/rollout_004/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_004/agentview/0040.png", "metadata/rollout_004/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_004/agentview/0041.png", "metadata/rollout_004/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_004/agentview/0042.png", "metadata/rollout_004/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0043.png", "metadata/rollout_004/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0044.png", "metadata/rollout_004/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0045.png", "metadata/rollout_004/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_004/agentview/0046.png", "metadata/rollout_004/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_004/agentview/0047.png", "metadata/rollout_004/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0048.png", "metadata/rollout_004/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_004/agentview/0049.png", "metadata/rollout_004/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_004/agentview/0049.png", "metadata/rollout_004/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0000.png", "metadata/rollout_005/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0001.png", "metadata/rollout_005/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0002.png", "metadata/rollout_005/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0003.png", "metadata/rollout_005/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0004.png", "metadata/rollout_005/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0005.png", "metadata/rollout_005/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0006.png", "metadata/rollout_005/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0007.png", "metadata/rollout_005/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0008.png", "metadata/rollout_005/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_005/agentview/0009.png", "metadata/rollout_005/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_005/agentview/0010.png", "metadata/rollout_005/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_005/agentview/0011.png", "metadata/rollout_005/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_005/agentview/0012.png", "metadata/rollout_005/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_005/agentview/0013.png", "metadata/rollout_005/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_005/agentview/0014.png", "metadata/rollout_005/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_005/agentview/0015.png", "metadata/rollout_005/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_005/agentview/0016.png", "metadata/rollout_005/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_005/agentview/0017.png", "metadata/rollout_005/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_005/agentview/0018.png", "metadata/rollout_005/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_005/agentview/0019.png", "metadata/rollout_005/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0020.png", "metadata/rollout_005/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0021.png", "metadata/rollout_005/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0022.png", "metadata/rollout_005/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0023.png", "metadata/rollout_005/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0024.png", "metadata/rollout_005/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0025.png", "metadata/rollout_005/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_005/agentview/0026.png", "metadata/rollout_005/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_005/agentview/0027.png", "metadata/rollout_005/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0028.png", "metadata/rollout_005/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0029.png", "metadata/rollout_005/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0030.png", "metadata/rollout_005/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0031.png", "metadata/rollout_005/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0032.png", "metadata/rollout_005/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0033.png", "metadata/rollout_005/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0034.png", "metadata/rollout_005/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0035.png", "metadata/rollout_005/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_005/agentview/0036.png", "metadata/rollout_005/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0037.png", "metadata/rollout_005/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0038.png", "metadata/rollout_005/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0039.png", "metadata/rollout_005/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0040.png", "metadata/rollout_005/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0041.png", "metadata/rollout_005/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0042.png", "metadata/rollout_005/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0043.png", "metadata/rollout_005/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0044.png", "metadata/rollout_005/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0045.png", "metadata/rollout_005/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0046.png", "metadata/rollout_005/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0047.png", "metadata/rollout_005/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0048.png", "metadata/rollout_005/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0049.png", "metadata/rollout_005/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0050.png", "metadata/rollout_005/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_005/agentview/0051.png", "metadata/rollout_005/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0052.png", "metadata/rollout_005/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0053.png", "metadata/rollout_005/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0054.png", "metadata/rollout_005/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_005/agentview/0055.png", "metadata/rollout_005/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_005/agentview/0056.png", "metadata/rollout_005/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0057.png", "metadata/rollout_005/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_005/agentview/0058.png", "metadata/rollout_005/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_005/agentview/0058.png", "metadata/rollout_005/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0000.png", "metadata/rollout_006/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0001.png", "metadata/rollout_006/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0002.png", "metadata/rollout_006/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0003.png", "metadata/rollout_006/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0004.png", "metadata/rollout_006/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0005.png", "metadata/rollout_006/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0006.png", "metadata/rollout_006/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0007.png", "metadata/rollout_006/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0008.png", "metadata/rollout_006/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_006/agentview/0009.png", "metadata/rollout_006/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_006/agentview/0010.png", "metadata/rollout_006/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_006/agentview/0011.png", "metadata/rollout_006/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_006/agentview/0012.png", "metadata/rollout_006/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_006/agentview/0013.png", "metadata/rollout_006/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_006/agentview/0014.png", "metadata/rollout_006/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_006/agentview/0015.png", "metadata/rollout_006/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_006/agentview/0016.png", "metadata/rollout_006/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0017.png", "metadata/rollout_006/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0018.png", "metadata/rollout_006/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0019.png", "metadata/rollout_006/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0020.png", "metadata/rollout_006/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0021.png", "metadata/rollout_006/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0022.png", "metadata/rollout_006/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_006/agentview/0023.png", "metadata/rollout_006/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_006/agentview/0024.png", "metadata/rollout_006/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0025.png", "metadata/rollout_006/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0026.png", "metadata/rollout_006/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0027.png", "metadata/rollout_006/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0028.png", "metadata/rollout_006/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0029.png", "metadata/rollout_006/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0030.png", "metadata/rollout_006/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0031.png", "metadata/rollout_006/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0032.png", "metadata/rollout_006/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_006/agentview/0033.png", "metadata/rollout_006/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_006/agentview/0034.png", "metadata/rollout_006/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_006/agentview/0035.png", "metadata/rollout_006/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_006/agentview/0036.png", "metadata/rollout_006/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_006/agentview/0037.png", "metadata/rollout_006/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_006/agentview/0038.png", "metadata/rollout_006/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_006/agentview/0039.png", "metadata/rollout_006/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_006/agentview/0040.png", "metadata/rollout_006/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_006/agentview/0041.png", "metadata/rollout_006/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_006/agentview/0042.png", "metadata/rollout_006/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0043.png", "metadata/rollout_006/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0044.png", "metadata/rollout_006/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0045.png", "metadata/rollout_006/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_006/agentview/0046.png", "metadata/rollout_006/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_006/agentview/0047.png", "metadata/rollout_006/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0048.png", "metadata/rollout_006/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_006/agentview/0049.png", "metadata/rollout_006/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_006/agentview/0049.png", "metadata/rollout_006/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0000.png", "metadata/rollout_007/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0001.png", "metadata/rollout_007/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0002.png", "metadata/rollout_007/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0003.png", "metadata/rollout_007/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0004.png", "metadata/rollout_007/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0005.png", "metadata/rollout_007/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0006.png", "metadata/rollout_007/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0007.png", "metadata/rollout_007/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0008.png", "metadata/rollout_007/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0009.png", "metadata/rollout_007/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_007/agentview/0010.png", "metadata/rollout_007/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0011.png", "metadata/rollout_007/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0012.png", "metadata/rollout_007/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0013.png", "metadata/rollout_007/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0014.png", "metadata/rollout_007/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0015.png", "metadata/rollout_007/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0016.png", "metadata/rollout_007/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0017.png", "metadata/rollout_007/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0018.png", "metadata/rollout_007/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0019.png", "metadata/rollout_007/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0020.png", "metadata/rollout_007/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0021.png", "metadata/rollout_007/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0022.png", "metadata/rollout_007/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0023.png", "metadata/rollout_007/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0024.png", "metadata/rollout_007/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0025.png", "metadata/rollout_007/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_007/agentview/0026.png", "metadata/rollout_007/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0027.png", "metadata/rollout_007/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0028.png", "metadata/rollout_007/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0029.png", "metadata/rollout_007/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0030.png", "metadata/rollout_007/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0031.png", "metadata/rollout_007/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0032.png", "metadata/rollout_007/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0033.png", "metadata/rollout_007/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0034.png", "metadata/rollout_007/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_007/agentview/0035.png", "metadata/rollout_007/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_007/agentview/0036.png", "metadata/rollout_007/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_007/agentview/0037.png", "metadata/rollout_007/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_007/agentview/0038.png", "metadata/rollout_007/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_007/agentview/0039.png", "metadata/rollout_007/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_007/agentview/0040.png", "metadata/rollout_007/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_007/agentview/0041.png", "metadata/rollout_007/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_007/agentview/0042.png", "metadata/rollout_007/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_007/agentview/0043.png", "metadata/rollout_007/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_007/agentview/0044.png", "metadata/rollout_007/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_007/agentview/0045.png", "metadata/rollout_007/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0046.png", "metadata/rollout_007/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0047.png", "metadata/rollout_007/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0048.png", "metadata/rollout_007/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_007/agentview/0049.png", "metadata/rollout_007/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_007/agentview/0050.png", "metadata/rollout_007/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0051.png", "metadata/rollout_007/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_007/agentview/0052.png", "metadata/rollout_007/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_007/agentview/0052.png", "metadata/rollout_007/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0000.png", "metadata/rollout_008/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0001.png", "metadata/rollout_008/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0002.png", "metadata/rollout_008/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0003.png", "metadata/rollout_008/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0004.png", "metadata/rollout_008/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0005.png", "metadata/rollout_008/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0006.png", "metadata/rollout_008/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0007.png", "metadata/rollout_008/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0008.png", "metadata/rollout_008/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_008/agentview/0009.png", "metadata/rollout_008/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0010.png", "metadata/rollout_008/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0011.png", "metadata/rollout_008/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0012.png", "metadata/rollout_008/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0013.png", "metadata/rollout_008/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0014.png", "metadata/rollout_008/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0015.png", "metadata/rollout_008/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0016.png", "metadata/rollout_008/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0017.png", "metadata/rollout_008/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0018.png", "metadata/rollout_008/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0019.png", "metadata/rollout_008/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0020.png", "metadata/rollout_008/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0021.png", "metadata/rollout_008/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0022.png", "metadata/rollout_008/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0023.png", "metadata/rollout_008/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0024.png", "metadata/rollout_008/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0025.png", "metadata/rollout_008/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0026.png", "metadata/rollout_008/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0027.png", "metadata/rollout_008/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0028.png", "metadata/rollout_008/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_008/agentview/0029.png", "metadata/rollout_008/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0030.png", "metadata/rollout_008/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0031.png", "metadata/rollout_008/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0032.png", "metadata/rollout_008/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0033.png", "metadata/rollout_008/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0034.png", "metadata/rollout_008/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0035.png", "metadata/rollout_008/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0036.png", "metadata/rollout_008/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0037.png", "metadata/rollout_008/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0038.png", "metadata/rollout_008/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0039.png", "metadata/rollout_008/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0040.png", "metadata/rollout_008/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0041.png", "metadata/rollout_008/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0042.png", "metadata/rollout_008/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0043.png", "metadata/rollout_008/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0044.png", "metadata/rollout_008/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0045.png", "metadata/rollout_008/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_008/agentview/0046.png", "metadata/rollout_008/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_008/agentview/0047.png", "metadata/rollout_008/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_008/agentview/0048.png", "metadata/rollout_008/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0049.png", "metadata/rollout_008/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0050.png", "metadata/rollout_008/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0051.png", "metadata/rollout_008/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_008/agentview/0052.png", "metadata/rollout_008/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_008/agentview/0053.png", "metadata/rollout_008/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0054.png", "metadata/rollout_008/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_008/agentview/0055.png", "metadata/rollout_008/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_008/agentview/0055.png", "metadata/rollout_008/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0000.png", "metadata/rollout_009/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0001.png", "metadata/rollout_009/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0002.png", "metadata/rollout_009/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0003.png", "metadata/rollout_009/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0004.png", "metadata/rollout_009/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0005.png", "metadata/rollout_009/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0006.png", "metadata/rollout_009/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0007.png", "metadata/rollout_009/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0008.png", "metadata/rollout_009/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_009/agentview/0009.png", "metadata/rollout_009/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0010.png", "metadata/rollout_009/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0011.png", "metadata/rollout_009/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0012.png", "metadata/rollout_009/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0013.png", "metadata/rollout_009/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0014.png", "metadata/rollout_009/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0015.png", "metadata/rollout_009/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0016.png", "metadata/rollout_009/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0017.png", "metadata/rollout_009/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0018.png", "metadata/rollout_009/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_009/agentview/0019.png", "metadata/rollout_009/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_009/agentview/0020.png", "metadata/rollout_009/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0021.png", "metadata/rollout_009/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0022.png", "metadata/rollout_009/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0023.png", "metadata/rollout_009/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0024.png", "metadata/rollout_009/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0025.png", "metadata/rollout_009/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0026.png", "metadata/rollout_009/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_009/agentview/0027.png", "metadata/rollout_009/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0028.png", "metadata/rollout_009/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0029.png", "metadata/rollout_009/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0030.png", "metadata/rollout_009/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0031.png", "metadata/rollout_009/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0032.png", "metadata/rollout_009/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0033.png", "metadata/rollout_009/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0034.png", "metadata/rollout_009/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0035.png", "metadata/rollout_009/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0036.png", "metadata/rollout_009/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0037.png", "metadata/rollout_009/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0038.png", "metadata/rollout_009/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0039.png", "metadata/rollout_009/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0040.png", "metadata/rollout_009/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0041.png", "metadata/rollout_009/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0042.png", "metadata/rollout_009/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0043.png", "metadata/rollout_009/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0044.png", "metadata/rollout_009/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_009/agentview/0045.png", "metadata/rollout_009/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0046.png", "metadata/rollout_009/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0047.png", "metadata/rollout_009/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0048.png", "metadata/rollout_009/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_009/agentview/0049.png", "metadata/rollout_009/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_009/agentview/0050.png", "metadata/rollout_009/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0051.png", "metadata/rollout_009/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_009/agentview/0052.png", "metadata/rollout_009/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the left bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_009/agentview/0052.png", "metadata/rollout_009/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0000.png", "metadata/rollout_010/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0001.png", "metadata/rollout_010/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0002.png", "metadata/rollout_010/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0003.png", "metadata/rollout_010/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0004.png", "metadata/rollout_010/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0005.png", "metadata/rollout_010/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0006.png", "metadata/rollout_010/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0007.png", "metadata/rollout_010/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0008.png", "metadata/rollout_010/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_010/agentview/0009.png", "metadata/rollout_010/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_010/agentview/0010.png", "metadata/rollout_010/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_010/agentview/0011.png", "metadata/rollout_010/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_010/agentview/0012.png", "metadata/rollout_010/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_010/agentview/0013.png", "metadata/rollout_010/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0014.png", "metadata/rollout_010/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0015.png", "metadata/rollout_010/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0016.png", "metadata/rollout_010/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0017.png", "metadata/rollout_010/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0018.png", "metadata/rollout_010/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0019.png", "metadata/rollout_010/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0020.png", "metadata/rollout_010/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0021.png", "metadata/rollout_010/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0022.png", "metadata/rollout_010/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_010/agentview/0023.png", "metadata/rollout_010/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0024.png", "metadata/rollout_010/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0025.png", "metadata/rollout_010/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0026.png", "metadata/rollout_010/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0027.png", "metadata/rollout_010/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0028.png", "metadata/rollout_010/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0029.png", "metadata/rollout_010/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0030.png", "metadata/rollout_010/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0031.png", "metadata/rollout_010/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0032.png", "metadata/rollout_010/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0033.png", "metadata/rollout_010/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0034.png", "metadata/rollout_010/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0035.png", "metadata/rollout_010/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0036.png", "metadata/rollout_010/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0037.png", "metadata/rollout_010/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0038.png", "metadata/rollout_010/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_010/agentview/0039.png", "metadata/rollout_010/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_010/agentview/0040.png", "metadata/rollout_010/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0041.png", "metadata/rollout_010/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0042.png", "metadata/rollout_010/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0043.png", "metadata/rollout_010/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_010/agentview/0044.png", "metadata/rollout_010/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_010/agentview/0045.png", "metadata/rollout_010/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0046.png", "metadata/rollout_010/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_010/agentview/0047.png", "metadata/rollout_010/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_010/agentview/0047.png", "metadata/rollout_010/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0000.png", "metadata/rollout_011/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0001.png", "metadata/rollout_011/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0002.png", "metadata/rollout_011/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0003.png", "metadata/rollout_011/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0004.png", "metadata/rollout_011/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0005.png", "metadata/rollout_011/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0006.png", "metadata/rollout_011/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0007.png", "metadata/rollout_011/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0008.png", "metadata/rollout_011/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0009.png", "metadata/rollout_011/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_011/agentview/0010.png", "metadata/rollout_011/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0011.png", "metadata/rollout_011/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0012.png", "metadata/rollout_011/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0013.png", "metadata/rollout_011/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0014.png", "metadata/rollout_011/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0015.png", "metadata/rollout_011/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0016.png", "metadata/rollout_011/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0017.png", "metadata/rollout_011/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0018.png", "metadata/rollout_011/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0019.png", "metadata/rollout_011/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_011/agentview/0020.png", "metadata/rollout_011/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_011/agentview/0021.png", "metadata/rollout_011/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_011/agentview/0022.png", "metadata/rollout_011/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0023.png", "metadata/rollout_011/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0024.png", "metadata/rollout_011/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0025.png", "metadata/rollout_011/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0026.png", "metadata/rollout_011/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0027.png", "metadata/rollout_011/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_011/agentview/0028.png", "metadata/rollout_011/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_011/agentview/0029.png", "metadata/rollout_011/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_011/agentview/0030.png", "metadata/rollout_011/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_011/agentview/0031.png", "metadata/rollout_011/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_011/agentview/0032.png", "metadata/rollout_011/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_011/agentview/0033.png", "metadata/rollout_011/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_011/agentview/0034.png", "metadata/rollout_011/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_011/agentview/0035.png", "metadata/rollout_011/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_011/agentview/0036.png", "metadata/rollout_011/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_011/agentview/0037.png", "metadata/rollout_011/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_011/agentview/0038.png", "metadata/rollout_011/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_011/agentview/0039.png", "metadata/rollout_011/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_011/agentview/0040.png", "metadata/rollout_011/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_011/agentview/0041.png", "metadata/rollout_011/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_011/agentview/0042.png", "metadata/rollout_011/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_011/agentview/0043.png", "metadata/rollout_011/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_011/agentview/0044.png", "metadata/rollout_011/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0045.png", "metadata/rollout_011/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0046.png", "metadata/rollout_011/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_011/agentview/0047.png", "metadata/rollout_011/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_011/agentview/0048.png", "metadata/rollout_011/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_011/agentview/0049.png", "metadata/rollout_011/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_011/agentview/0050.png", "metadata/rollout_011/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_011/agentview/0050.png", "metadata/rollout_011/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0000.png", "metadata/rollout_012/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0001.png", "metadata/rollout_012/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0002.png", "metadata/rollout_012/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0003.png", "metadata/rollout_012/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0004.png", "metadata/rollout_012/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0005.png", "metadata/rollout_012/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0006.png", "metadata/rollout_012/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0007.png", "metadata/rollout_012/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0008.png", "metadata/rollout_012/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_012/agentview/0009.png", "metadata/rollout_012/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_012/agentview/0010.png", "metadata/rollout_012/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_012/agentview/0011.png", "metadata/rollout_012/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_012/agentview/0012.png", "metadata/rollout_012/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_012/agentview/0013.png", "metadata/rollout_012/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0014.png", "metadata/rollout_012/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0015.png", "metadata/rollout_012/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0016.png", "metadata/rollout_012/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0017.png", "metadata/rollout_012/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0018.png", "metadata/rollout_012/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0019.png", "metadata/rollout_012/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_012/agentview/0020.png", "metadata/rollout_012/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0021.png", "metadata/rollout_012/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0022.png", "metadata/rollout_012/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0023.png", "metadata/rollout_012/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0024.png", "metadata/rollout_012/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0025.png", "metadata/rollout_012/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0026.png", "metadata/rollout_012/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0027.png", "metadata/rollout_012/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0028.png", "metadata/rollout_012/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_012/agentview/0029.png", "metadata/rollout_012/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_012/agentview/0030.png", "metadata/rollout_012/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_012/agentview/0031.png", "metadata/rollout_012/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0032.png", "metadata/rollout_012/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0033.png", "metadata/rollout_012/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0034.png", "metadata/rollout_012/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0035.png", "metadata/rollout_012/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0036.png", "metadata/rollout_012/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0037.png", "metadata/rollout_012/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0038.png", "metadata/rollout_012/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0039.png", "metadata/rollout_012/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0040.png", "metadata/rollout_012/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0041.png", "metadata/rollout_012/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_012/agentview/0042.png", "metadata/rollout_012/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0043.png", "metadata/rollout_012/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0044.png", "metadata/rollout_012/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0045.png", "metadata/rollout_012/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_012/agentview/0046.png", "metadata/rollout_012/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_012/agentview/0047.png", "metadata/rollout_012/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0048.png", "metadata/rollout_012/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_012/agentview/0049.png", "metadata/rollout_012/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_012/agentview/0049.png", "metadata/rollout_012/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0000.png", "metadata/rollout_013/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0001.png", "metadata/rollout_013/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0002.png", "metadata/rollout_013/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0003.png", "metadata/rollout_013/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0004.png", "metadata/rollout_013/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0005.png", "metadata/rollout_013/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0006.png", "metadata/rollout_013/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0007.png", "metadata/rollout_013/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0008.png", "metadata/rollout_013/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0009.png", "metadata/rollout_013/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0010.png", "metadata/rollout_013/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0011.png", "metadata/rollout_013/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0012.png", "metadata/rollout_013/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0013.png", "metadata/rollout_013/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0014.png", "metadata/rollout_013/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0015.png", "metadata/rollout_013/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0016.png", "metadata/rollout_013/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0017.png", "metadata/rollout_013/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_013/agentview/0018.png", "metadata/rollout_013/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_013/agentview/0019.png", "metadata/rollout_013/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0020.png", "metadata/rollout_013/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0021.png", "metadata/rollout_013/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0022.png", "metadata/rollout_013/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0023.png", "metadata/rollout_013/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0024.png", "metadata/rollout_013/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0025.png", "metadata/rollout_013/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_013/agentview/0026.png", "metadata/rollout_013/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0027.png", "metadata/rollout_013/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0028.png", "metadata/rollout_013/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0029.png", "metadata/rollout_013/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0030.png", "metadata/rollout_013/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0031.png", "metadata/rollout_013/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0032.png", "metadata/rollout_013/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0033.png", "metadata/rollout_013/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0034.png", "metadata/rollout_013/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0035.png", "metadata/rollout_013/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0036.png", "metadata/rollout_013/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0037.png", "metadata/rollout_013/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0038.png", "metadata/rollout_013/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0039.png", "metadata/rollout_013/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0040.png", "metadata/rollout_013/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0041.png", "metadata/rollout_013/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0042.png", "metadata/rollout_013/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0043.png", "metadata/rollout_013/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0044.png", "metadata/rollout_013/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_013/agentview/0045.png", "metadata/rollout_013/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_013/agentview/0046.png", "metadata/rollout_013/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0047.png", "metadata/rollout_013/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0048.png", "metadata/rollout_013/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0049.png", "metadata/rollout_013/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_013/agentview/0050.png", "metadata/rollout_013/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_013/agentview/0051.png", "metadata/rollout_013/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0052.png", "metadata/rollout_013/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_013/agentview/0053.png", "metadata/rollout_013/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_013/agentview/0053.png", "metadata/rollout_013/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0000.png", "metadata/rollout_014/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0001.png", "metadata/rollout_014/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0002.png", "metadata/rollout_014/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0003.png", "metadata/rollout_014/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0004.png", "metadata/rollout_014/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0005.png", "metadata/rollout_014/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0006.png", "metadata/rollout_014/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0007.png", "metadata/rollout_014/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0008.png", "metadata/rollout_014/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0009.png", "metadata/rollout_014/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0010.png", "metadata/rollout_014/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0011.png", "metadata/rollout_014/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0012.png", "metadata/rollout_014/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0013.png", "metadata/rollout_014/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0014.png", "metadata/rollout_014/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0015.png", "metadata/rollout_014/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0016.png", "metadata/rollout_014/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0017.png", "metadata/rollout_014/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0018.png", "metadata/rollout_014/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0019.png", "metadata/rollout_014/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0020.png", "metadata/rollout_014/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0021.png", "metadata/rollout_014/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0022.png", "metadata/rollout_014/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0023.png", "metadata/rollout_014/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0024.png", "metadata/rollout_014/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0025.png", "metadata/rollout_014/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_014/agentview/0026.png", "metadata/rollout_014/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0027.png", "metadata/rollout_014/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0028.png", "metadata/rollout_014/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0029.png", "metadata/rollout_014/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0030.png", "metadata/rollout_014/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0031.png", "metadata/rollout_014/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0032.png", "metadata/rollout_014/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0033.png", "metadata/rollout_014/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0034.png", "metadata/rollout_014/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0035.png", "metadata/rollout_014/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0036.png", "metadata/rollout_014/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0037.png", "metadata/rollout_014/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0038.png", "metadata/rollout_014/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0039.png", "metadata/rollout_014/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0040.png", "metadata/rollout_014/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0041.png", "metadata/rollout_014/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_014/agentview/0042.png", "metadata/rollout_014/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_014/agentview/0043.png", "metadata/rollout_014/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0044.png", "metadata/rollout_014/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0045.png", "metadata/rollout_014/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0046.png", "metadata/rollout_014/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_014/agentview/0047.png", "metadata/rollout_014/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_014/agentview/0048.png", "metadata/rollout_014/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0049.png", "metadata/rollout_014/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_014/agentview/0050.png", "metadata/rollout_014/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_014/agentview/0050.png", "metadata/rollout_014/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0000.png", "metadata/rollout_015/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0001.png", "metadata/rollout_015/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0002.png", "metadata/rollout_015/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0003.png", "metadata/rollout_015/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0004.png", "metadata/rollout_015/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0005.png", "metadata/rollout_015/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0006.png", "metadata/rollout_015/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0007.png", "metadata/rollout_015/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0008.png", "metadata/rollout_015/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0009.png", "metadata/rollout_015/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0010.png", "metadata/rollout_015/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0011.png", "metadata/rollout_015/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0012.png", "metadata/rollout_015/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_015/agentview/0013.png", "metadata/rollout_015/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_015/agentview/0014.png", "metadata/rollout_015/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_015/agentview/0015.png", "metadata/rollout_015/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_015/agentview/0016.png", "metadata/rollout_015/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0017.png", "metadata/rollout_015/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0018.png", "metadata/rollout_015/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0019.png", "metadata/rollout_015/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0020.png", "metadata/rollout_015/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0021.png", "metadata/rollout_015/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0022.png", "metadata/rollout_015/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_015/agentview/0023.png", "metadata/rollout_015/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0024.png", "metadata/rollout_015/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0025.png", "metadata/rollout_015/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0026.png", "metadata/rollout_015/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0027.png", "metadata/rollout_015/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0028.png", "metadata/rollout_015/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0029.png", "metadata/rollout_015/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0030.png", "metadata/rollout_015/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0031.png", "metadata/rollout_015/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0032.png", "metadata/rollout_015/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0033.png", "metadata/rollout_015/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0034.png", "metadata/rollout_015/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0035.png", "metadata/rollout_015/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0036.png", "metadata/rollout_015/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0037.png", "metadata/rollout_015/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0038.png", "metadata/rollout_015/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_015/agentview/0039.png", "metadata/rollout_015/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0040.png", "metadata/rollout_015/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0041.png", "metadata/rollout_015/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0042.png", "metadata/rollout_015/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_015/agentview/0043.png", "metadata/rollout_015/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_015/agentview/0044.png", "metadata/rollout_015/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0045.png", "metadata/rollout_015/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_015/agentview/0046.png", "metadata/rollout_015/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_015/agentview/0046.png", "metadata/rollout_015/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0000.png", "metadata/rollout_016/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0001.png", "metadata/rollout_016/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0002.png", "metadata/rollout_016/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0003.png", "metadata/rollout_016/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0004.png", "metadata/rollout_016/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0005.png", "metadata/rollout_016/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0006.png", "metadata/rollout_016/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0007.png", "metadata/rollout_016/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0008.png", "metadata/rollout_016/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0009.png", "metadata/rollout_016/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_016/agentview/0010.png", "metadata/rollout_016/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0011.png", "metadata/rollout_016/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0012.png", "metadata/rollout_016/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0013.png", "metadata/rollout_016/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0014.png", "metadata/rollout_016/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0015.png", "metadata/rollout_016/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0016.png", "metadata/rollout_016/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0017.png", "metadata/rollout_016/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0018.png", "metadata/rollout_016/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0019.png", "metadata/rollout_016/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0020.png", "metadata/rollout_016/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0021.png", "metadata/rollout_016/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0022.png", "metadata/rollout_016/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0023.png", "metadata/rollout_016/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0024.png", "metadata/rollout_016/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0025.png", "metadata/rollout_016/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_016/agentview/0026.png", "metadata/rollout_016/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0027.png", "metadata/rollout_016/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0028.png", "metadata/rollout_016/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0029.png", "metadata/rollout_016/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0030.png", "metadata/rollout_016/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0031.png", "metadata/rollout_016/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0032.png", "metadata/rollout_016/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0033.png", "metadata/rollout_016/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0034.png", "metadata/rollout_016/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0035.png", "metadata/rollout_016/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0036.png", "metadata/rollout_016/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0037.png", "metadata/rollout_016/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0038.png", "metadata/rollout_016/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0039.png", "metadata/rollout_016/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0040.png", "metadata/rollout_016/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0041.png", "metadata/rollout_016/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0042.png", "metadata/rollout_016/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0043.png", "metadata/rollout_016/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0044.png", "metadata/rollout_016/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_016/agentview/0045.png", "metadata/rollout_016/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_016/agentview/0046.png", "metadata/rollout_016/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_016/agentview/0047.png", "metadata/rollout_016/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_016/agentview/0048.png", "metadata/rollout_016/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_016/agentview/0049.png", "metadata/rollout_016/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_016/agentview/0050.png", "metadata/rollout_016/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0051.png", "metadata/rollout_016/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0052.png", "metadata/rollout_016/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0053.png", "metadata/rollout_016/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_016/agentview/0054.png", "metadata/rollout_016/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_016/agentview/0055.png", "metadata/rollout_016/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0056.png", "metadata/rollout_016/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_016/agentview/0057.png", "metadata/rollout_016/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_016/agentview/0057.png", "metadata/rollout_016/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0000.png", "metadata/rollout_017/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0001.png", "metadata/rollout_017/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0002.png", "metadata/rollout_017/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0003.png", "metadata/rollout_017/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0004.png", "metadata/rollout_017/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0005.png", "metadata/rollout_017/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0006.png", "metadata/rollout_017/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0007.png", "metadata/rollout_017/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0008.png", "metadata/rollout_017/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_017/agentview/0009.png", "metadata/rollout_017/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_017/agentview/0010.png", "metadata/rollout_017/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_017/agentview/0011.png", "metadata/rollout_017/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_017/agentview/0012.png", "metadata/rollout_017/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0013.png", "metadata/rollout_017/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0014.png", "metadata/rollout_017/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0015.png", "metadata/rollout_017/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0016.png", "metadata/rollout_017/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0017.png", "metadata/rollout_017/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0018.png", "metadata/rollout_017/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0019.png", "metadata/rollout_017/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0020.png", "metadata/rollout_017/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0021.png", "metadata/rollout_017/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_017/agentview/0022.png", "metadata/rollout_017/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0023.png", "metadata/rollout_017/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0024.png", "metadata/rollout_017/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0025.png", "metadata/rollout_017/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0026.png", "metadata/rollout_017/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0027.png", "metadata/rollout_017/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0028.png", "metadata/rollout_017/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0029.png", "metadata/rollout_017/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0030.png", "metadata/rollout_017/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0031.png", "metadata/rollout_017/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0032.png", "metadata/rollout_017/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0033.png", "metadata/rollout_017/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0034.png", "metadata/rollout_017/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0035.png", "metadata/rollout_017/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0036.png", "metadata/rollout_017/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0037.png", "metadata/rollout_017/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0038.png", "metadata/rollout_017/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_017/agentview/0039.png", "metadata/rollout_017/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_017/agentview/0040.png", "metadata/rollout_017/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_017/agentview/0041.png", "metadata/rollout_017/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_017/agentview/0042.png", "metadata/rollout_017/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_017/agentview/0043.png", "metadata/rollout_017/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_017/agentview/0044.png", "metadata/rollout_017/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0045.png", "metadata/rollout_017/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0046.png", "metadata/rollout_017/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0047.png", "metadata/rollout_017/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_017/agentview/0048.png", "metadata/rollout_017/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_017/agentview/0049.png", "metadata/rollout_017/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0050.png", "metadata/rollout_017/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_017/agentview/0051.png", "metadata/rollout_017/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_017/agentview/0051.png", "metadata/rollout_017/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0000.png", "metadata/rollout_018/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0001.png", "metadata/rollout_018/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0002.png", "metadata/rollout_018/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0003.png", "metadata/rollout_018/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0004.png", "metadata/rollout_018/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0005.png", "metadata/rollout_018/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0006.png", "metadata/rollout_018/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0007.png", "metadata/rollout_018/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0008.png", "metadata/rollout_018/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_018/agentview/0009.png", "metadata/rollout_018/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0010.png", "metadata/rollout_018/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0011.png", "metadata/rollout_018/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0012.png", "metadata/rollout_018/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0013.png", "metadata/rollout_018/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0014.png", "metadata/rollout_018/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0015.png", "metadata/rollout_018/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0016.png", "metadata/rollout_018/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0017.png", "metadata/rollout_018/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0018.png", "metadata/rollout_018/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0019.png", "metadata/rollout_018/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0020.png", "metadata/rollout_018/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0021.png", "metadata/rollout_018/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0022.png", "metadata/rollout_018/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0023.png", "metadata/rollout_018/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0024.png", "metadata/rollout_018/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0025.png", "metadata/rollout_018/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0026.png", "metadata/rollout_018/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_018/agentview/0027.png", "metadata/rollout_018/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0028.png", "metadata/rollout_018/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0029.png", "metadata/rollout_018/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0030.png", "metadata/rollout_018/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0031.png", "metadata/rollout_018/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0032.png", "metadata/rollout_018/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0033.png", "metadata/rollout_018/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0034.png", "metadata/rollout_018/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0035.png", "metadata/rollout_018/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0036.png", "metadata/rollout_018/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0037.png", "metadata/rollout_018/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0038.png", "metadata/rollout_018/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0039.png", "metadata/rollout_018/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0040.png", "metadata/rollout_018/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0041.png", "metadata/rollout_018/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0042.png", "metadata/rollout_018/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_018/agentview/0043.png", "metadata/rollout_018/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0044.png", "metadata/rollout_018/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0045.png", "metadata/rollout_018/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0046.png", "metadata/rollout_018/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_018/agentview/0047.png", "metadata/rollout_018/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_018/agentview/0048.png", "metadata/rollout_018/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0049.png", "metadata/rollout_018/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_018/agentview/0050.png", "metadata/rollout_018/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_018/agentview/0050.png", "metadata/rollout_018/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0000.png", "metadata/rollout_019/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0001.png", "metadata/rollout_019/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0002.png", "metadata/rollout_019/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0003.png", "metadata/rollout_019/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0004.png", "metadata/rollout_019/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0005.png", "metadata/rollout_019/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0006.png", "metadata/rollout_019/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0007.png", "metadata/rollout_019/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0008.png", "metadata/rollout_019/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_019/agentview/0009.png", "metadata/rollout_019/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0010.png", "metadata/rollout_019/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0011.png", "metadata/rollout_019/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0012.png", "metadata/rollout_019/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0013.png", "metadata/rollout_019/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0014.png", "metadata/rollout_019/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0015.png", "metadata/rollout_019/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0016.png", "metadata/rollout_019/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0017.png", "metadata/rollout_019/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0018.png", "metadata/rollout_019/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_019/agentview/0019.png", "metadata/rollout_019/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0020.png", "metadata/rollout_019/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0021.png", "metadata/rollout_019/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0022.png", "metadata/rollout_019/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0023.png", "metadata/rollout_019/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0024.png", "metadata/rollout_019/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0025.png", "metadata/rollout_019/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_019/agentview/0026.png", "metadata/rollout_019/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0027.png", "metadata/rollout_019/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0028.png", "metadata/rollout_019/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0029.png", "metadata/rollout_019/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0030.png", "metadata/rollout_019/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0031.png", "metadata/rollout_019/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0032.png", "metadata/rollout_019/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0033.png", "metadata/rollout_019/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0034.png", "metadata/rollout_019/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_019/agentview/0035.png", "metadata/rollout_019/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_019/agentview/0036.png", "metadata/rollout_019/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_019/agentview/0037.png", "metadata/rollout_019/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_019/agentview/0038.png", "metadata/rollout_019/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_019/agentview/0039.png", "metadata/rollout_019/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_019/agentview/0040.png", "metadata/rollout_019/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_019/agentview/0041.png", "metadata/rollout_019/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_019/agentview/0042.png", "metadata/rollout_019/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_019/agentview/0043.png", "metadata/rollout_019/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0044.png", "metadata/rollout_019/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0045.png", "metadata/rollout_019/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0046.png", "metadata/rollout_019/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_019/agentview/0047.png", "metadata/rollout_019/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_019/agentview/0048.png", "metadata/rollout_019/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0049.png", "metadata/rollout_019/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_019/agentview/0050.png", "metadata/rollout_019/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the right bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_019/agentview/0050.png", "metadata/rollout_019/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0000.png", "metadata/rollout_020/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0001.png", "metadata/rollout_020/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0002.png", "metadata/rollout_020/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0003.png", "metadata/rollout_020/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0004.png", "metadata/rollout_020/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0005.png", "metadata/rollout_020/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0006.png", "metadata/rollout_020/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0007.png", "metadata/rollout_020/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0008.png", "metadata/rollout_020/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_020/agentview/0009.png", "metadata/rollout_020/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_020/agentview/0010.png", "metadata/rollout_020/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_020/agentview/0011.png", "metadata/rollout_020/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_020/agentview/0012.png", "metadata/rollout_020/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_020/agentview/0013.png", "metadata/rollout_020/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_020/agentview/0014.png", "metadata/rollout_020/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_020/agentview/0015.png", "metadata/rollout_020/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0016.png", "metadata/rollout_020/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0017.png", "metadata/rollout_020/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0018.png", "metadata/rollout_020/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0019.png", "metadata/rollout_020/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0020.png", "metadata/rollout_020/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0021.png", "metadata/rollout_020/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_020/agentview/0022.png", "metadata/rollout_020/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0023.png", "metadata/rollout_020/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0024.png", "metadata/rollout_020/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0025.png", "metadata/rollout_020/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0026.png", "metadata/rollout_020/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0027.png", "metadata/rollout_020/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0028.png", "metadata/rollout_020/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0029.png", "metadata/rollout_020/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0030.png", "metadata/rollout_020/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_020/agentview/0031.png", "metadata/rollout_020/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_020/agentview/0032.png", "metadata/rollout_020/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_020/agentview/0033.png", "metadata/rollout_020/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_020/agentview/0034.png", "metadata/rollout_020/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_020/agentview/0035.png", "metadata/rollout_020/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_020/agentview/0036.png", "metadata/rollout_020/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_020/agentview/0037.png", "metadata/rollout_020/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_020/agentview/0038.png", "metadata/rollout_020/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_020/agentview/0039.png", "metadata/rollout_020/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0040.png", "metadata/rollout_020/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0041.png", "metadata/rollout_020/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0042.png", "metadata/rollout_020/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_020/agentview/0043.png", "metadata/rollout_020/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_020/agentview/0044.png", "metadata/rollout_020/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0045.png", "metadata/rollout_020/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_020/agentview/0046.png", "metadata/rollout_020/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_020/agentview/0046.png", "metadata/rollout_020/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0000.png", "metadata/rollout_021/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0001.png", "metadata/rollout_021/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0002.png", "metadata/rollout_021/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0003.png", "metadata/rollout_021/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0004.png", "metadata/rollout_021/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0005.png", "metadata/rollout_021/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0006.png", "metadata/rollout_021/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0007.png", "metadata/rollout_021/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0008.png", "metadata/rollout_021/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_021/agentview/0009.png", "metadata/rollout_021/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_021/agentview/0010.png", "metadata/rollout_021/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_021/agentview/0011.png", "metadata/rollout_021/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_021/agentview/0012.png", "metadata/rollout_021/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_021/agentview/0013.png", "metadata/rollout_021/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_021/agentview/0014.png", "metadata/rollout_021/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_021/agentview/0015.png", "metadata/rollout_021/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_021/agentview/0016.png", "metadata/rollout_021/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_021/agentview/0017.png", "metadata/rollout_021/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0018.png", "metadata/rollout_021/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0019.png", "metadata/rollout_021/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0020.png", "metadata/rollout_021/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0021.png", "metadata/rollout_021/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0022.png", "metadata/rollout_021/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0023.png", "metadata/rollout_021/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_021/agentview/0024.png", "metadata/rollout_021/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0025.png", "metadata/rollout_021/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0026.png", "metadata/rollout_021/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0027.png", "metadata/rollout_021/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0028.png", "metadata/rollout_021/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0029.png", "metadata/rollout_021/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0030.png", "metadata/rollout_021/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0031.png", "metadata/rollout_021/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0032.png", "metadata/rollout_021/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0033.png", "metadata/rollout_021/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0034.png", "metadata/rollout_021/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0035.png", "metadata/rollout_021/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0036.png", "metadata/rollout_021/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0037.png", "metadata/rollout_021/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0038.png", "metadata/rollout_021/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0039.png", "metadata/rollout_021/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0040.png", "metadata/rollout_021/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0041.png", "metadata/rollout_021/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0042.png", "metadata/rollout_021/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_021/agentview/0043.png", "metadata/rollout_021/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_021/agentview/0044.png", "metadata/rollout_021/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_021/agentview/0045.png", "metadata/rollout_021/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_021/agentview/0046.png", "metadata/rollout_021/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0047.png", "metadata/rollout_021/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0048.png", "metadata/rollout_021/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0049.png", "metadata/rollout_021/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_021/agentview/0050.png", "metadata/rollout_021/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_021/agentview/0051.png", "metadata/rollout_021/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0052.png", "metadata/rollout_021/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_021/agentview/0053.png", "metadata/rollout_021/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_021/agentview/0053.png", "metadata/rollout_021/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0000.png", "metadata/rollout_022/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0001.png", "metadata/rollout_022/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0002.png", "metadata/rollout_022/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0003.png", "metadata/rollout_022/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0004.png", "metadata/rollout_022/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0005.png", "metadata/rollout_022/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0006.png", "metadata/rollout_022/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0007.png", "metadata/rollout_022/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0008.png", "metadata/rollout_022/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_022/agentview/0009.png", "metadata/rollout_022/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_022/agentview/0010.png", "metadata/rollout_022/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_022/agentview/0011.png", "metadata/rollout_022/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_022/agentview/0012.png", "metadata/rollout_022/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_022/agentview/0013.png", "metadata/rollout_022/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_022/agentview/0014.png", "metadata/rollout_022/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_022/agentview/0015.png", "metadata/rollout_022/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_022/agentview/0016.png", "metadata/rollout_022/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_022/agentview/0017.png", "metadata/rollout_022/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_022/agentview/0018.png", "metadata/rollout_022/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_022/agentview/0019.png", "metadata/rollout_022/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_022/agentview/0020.png", "metadata/rollout_022/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_022/agentview/0021.png", "metadata/rollout_022/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_022/agentview/0022.png", "metadata/rollout_022/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_022/agentview/0023.png", "metadata/rollout_022/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_022/agentview/0024.png", "metadata/rollout_022/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_022/agentview/0025.png", "metadata/rollout_022/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0026.png", "metadata/rollout_022/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0027.png", "metadata/rollout_022/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0028.png", "metadata/rollout_022/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0029.png", "metadata/rollout_022/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0030.png", "metadata/rollout_022/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0031.png", "metadata/rollout_022/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_022/agentview/0032.png", "metadata/rollout_022/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0033.png", "metadata/rollout_022/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0034.png", "metadata/rollout_022/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0035.png", "metadata/rollout_022/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0036.png", "metadata/rollout_022/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0037.png", "metadata/rollout_022/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0038.png", "metadata/rollout_022/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0039.png", "metadata/rollout_022/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0040.png", "metadata/rollout_022/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_022/agentview/0041.png", "metadata/rollout_022/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_022/agentview/0042.png", "metadata/rollout_022/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_022/agentview/0043.png", "metadata/rollout_022/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0044.png", "metadata/rollout_022/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0045.png", "metadata/rollout_022/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0046.png", "metadata/rollout_022/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0047.png", "metadata/rollout_022/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0048.png", "metadata/rollout_022/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0049.png", "metadata/rollout_022/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0050.png", "metadata/rollout_022/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0051.png", "metadata/rollout_022/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0052.png", "metadata/rollout_022/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0053.png", "metadata/rollout_022/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0054.png", "metadata/rollout_022/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0055.png", "metadata/rollout_022/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_022/agentview/0056.png", "metadata/rollout_022/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0057.png", "metadata/rollout_022/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0058.png", "metadata/rollout_022/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0059.png", "metadata/rollout_022/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_022/agentview/0060.png", "metadata/rollout_022/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_022/agentview/0061.png", "metadata/rollout_022/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0062.png", "metadata/rollout_022/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_022/agentview/0063.png", "metadata/rollout_022/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_022/agentview/0063.png", "metadata/rollout_022/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0000.png", "metadata/rollout_023/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0001.png", "metadata/rollout_023/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0002.png", "metadata/rollout_023/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0003.png", "metadata/rollout_023/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0004.png", "metadata/rollout_023/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0005.png", "metadata/rollout_023/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0006.png", "metadata/rollout_023/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0007.png", "metadata/rollout_023/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0008.png", "metadata/rollout_023/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0009.png", "metadata/rollout_023/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0010.png", "metadata/rollout_023/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0011.png", "metadata/rollout_023/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0012.png", "metadata/rollout_023/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0013.png", "metadata/rollout_023/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0014.png", "metadata/rollout_023/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0015.png", "metadata/rollout_023/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0016.png", "metadata/rollout_023/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0017.png", "metadata/rollout_023/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_023/agentview/0018.png", "metadata/rollout_023/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_023/agentview/0019.png", "metadata/rollout_023/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_023/agentview/0020.png", "metadata/rollout_023/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_023/agentview/0021.png", "metadata/rollout_023/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_023/agentview/0022.png", "metadata/rollout_023/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_023/agentview/0023.png", "metadata/rollout_023/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_023/agentview/0024.png", "metadata/rollout_023/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0025.png", "metadata/rollout_023/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0026.png", "metadata/rollout_023/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0027.png", "metadata/rollout_023/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0028.png", "metadata/rollout_023/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0029.png", "metadata/rollout_023/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0030.png", "metadata/rollout_023/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_023/agentview/0031.png", "metadata/rollout_023/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0032.png", "metadata/rollout_023/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0033.png", "metadata/rollout_023/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0034.png", "metadata/rollout_023/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0035.png", "metadata/rollout_023/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0036.png", "metadata/rollout_023/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0037.png", "metadata/rollout_023/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0038.png", "metadata/rollout_023/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0039.png", "metadata/rollout_023/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0040.png", "metadata/rollout_023/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0041.png", "metadata/rollout_023/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0042.png", "metadata/rollout_023/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0043.png", "metadata/rollout_023/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0044.png", "metadata/rollout_023/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0045.png", "metadata/rollout_023/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0046.png", "metadata/rollout_023/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0047.png", "metadata/rollout_023/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0048.png", "metadata/rollout_023/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0049.png", "metadata/rollout_023/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0050.png", "metadata/rollout_023/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0051.png", "metadata/rollout_023/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0052.png", "metadata/rollout_023/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_023/agentview/0053.png", "metadata/rollout_023/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_023/agentview/0054.png", "metadata/rollout_023/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0055.png", "metadata/rollout_023/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0056.png", "metadata/rollout_023/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0057.png", "metadata/rollout_023/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_023/agentview/0058.png", "metadata/rollout_023/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_023/agentview/0059.png", "metadata/rollout_023/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0060.png", "metadata/rollout_023/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_023/agentview/0061.png", "metadata/rollout_023/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_023/agentview/0061.png", "metadata/rollout_023/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0000.png", "metadata/rollout_024/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0001.png", "metadata/rollout_024/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0002.png", "metadata/rollout_024/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0003.png", "metadata/rollout_024/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0004.png", "metadata/rollout_024/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0005.png", "metadata/rollout_024/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0006.png", "metadata/rollout_024/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0007.png", "metadata/rollout_024/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0008.png", "metadata/rollout_024/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_024/agentview/0009.png", "metadata/rollout_024/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_024/agentview/0010.png", "metadata/rollout_024/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_024/agentview/0011.png", "metadata/rollout_024/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_024/agentview/0012.png", "metadata/rollout_024/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_024/agentview/0013.png", "metadata/rollout_024/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_024/agentview/0014.png", "metadata/rollout_024/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_024/agentview/0015.png", "metadata/rollout_024/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_024/agentview/0016.png", "metadata/rollout_024/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_024/agentview/0017.png", "metadata/rollout_024/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_024/agentview/0018.png", "metadata/rollout_024/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_024/agentview/0019.png", "metadata/rollout_024/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_024/agentview/0020.png", "metadata/rollout_024/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0021.png", "metadata/rollout_024/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0022.png", "metadata/rollout_024/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0023.png", "metadata/rollout_024/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0024.png", "metadata/rollout_024/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0025.png", "metadata/rollout_024/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0026.png", "metadata/rollout_024/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_024/agentview/0027.png", "metadata/rollout_024/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0028.png", "metadata/rollout_024/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0029.png", "metadata/rollout_024/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0030.png", "metadata/rollout_024/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0031.png", "metadata/rollout_024/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0032.png", "metadata/rollout_024/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0033.png", "metadata/rollout_024/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0034.png", "metadata/rollout_024/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0035.png", "metadata/rollout_024/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_024/agentview/0036.png", "metadata/rollout_024/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_024/agentview/0037.png", "metadata/rollout_024/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_024/agentview/0038.png", "metadata/rollout_024/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_024/agentview/0039.png", "metadata/rollout_024/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_024/agentview/0040.png", "metadata/rollout_024/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_024/agentview/0041.png", "metadata/rollout_024/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_024/agentview/0042.png", "metadata/rollout_024/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_024/agentview/0043.png", "metadata/rollout_024/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_024/agentview/0044.png", "metadata/rollout_024/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_024/agentview/0045.png", "metadata/rollout_024/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0046.png", "metadata/rollout_024/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0047.png", "metadata/rollout_024/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0048.png", "metadata/rollout_024/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_024/agentview/0049.png", "metadata/rollout_024/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_024/agentview/0050.png", "metadata/rollout_024/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0051.png", "metadata/rollout_024/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_024/agentview/0052.png", "metadata/rollout_024/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_024/agentview/0052.png", "metadata/rollout_024/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0000.png", "metadata/rollout_025/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0001.png", "metadata/rollout_025/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0002.png", "metadata/rollout_025/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0003.png", "metadata/rollout_025/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0004.png", "metadata/rollout_025/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0005.png", "metadata/rollout_025/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0006.png", "metadata/rollout_025/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0007.png", "metadata/rollout_025/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0008.png", "metadata/rollout_025/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_025/agentview/0009.png", "metadata/rollout_025/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_025/agentview/0010.png", "metadata/rollout_025/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_025/agentview/0011.png", "metadata/rollout_025/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_025/agentview/0012.png", "metadata/rollout_025/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_025/agentview/0013.png", "metadata/rollout_025/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0014.png", "metadata/rollout_025/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0015.png", "metadata/rollout_025/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0016.png", "metadata/rollout_025/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0017.png", "metadata/rollout_025/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0018.png", "metadata/rollout_025/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0019.png", "metadata/rollout_025/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_025/agentview/0020.png", "metadata/rollout_025/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0021.png", "metadata/rollout_025/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0022.png", "metadata/rollout_025/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0023.png", "metadata/rollout_025/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0024.png", "metadata/rollout_025/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0025.png", "metadata/rollout_025/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0026.png", "metadata/rollout_025/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0027.png", "metadata/rollout_025/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0028.png", "metadata/rollout_025/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0029.png", "metadata/rollout_025/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0030.png", "metadata/rollout_025/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0031.png", "metadata/rollout_025/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0032.png", "metadata/rollout_025/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0033.png", "metadata/rollout_025/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0034.png", "metadata/rollout_025/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0035.png", "metadata/rollout_025/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0036.png", "metadata/rollout_025/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0037.png", "metadata/rollout_025/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0038.png", "metadata/rollout_025/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_025/agentview/0039.png", "metadata/rollout_025/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0040.png", "metadata/rollout_025/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0041.png", "metadata/rollout_025/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0042.png", "metadata/rollout_025/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_025/agentview/0043.png", "metadata/rollout_025/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_025/agentview/0044.png", "metadata/rollout_025/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0045.png", "metadata/rollout_025/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_025/agentview/0046.png", "metadata/rollout_025/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_025/agentview/0046.png", "metadata/rollout_025/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0000.png", "metadata/rollout_026/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0001.png", "metadata/rollout_026/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0002.png", "metadata/rollout_026/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0003.png", "metadata/rollout_026/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0004.png", "metadata/rollout_026/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0005.png", "metadata/rollout_026/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0006.png", "metadata/rollout_026/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0007.png", "metadata/rollout_026/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0008.png", "metadata/rollout_026/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0009.png", "metadata/rollout_026/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_026/agentview/0010.png", "metadata/rollout_026/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0011.png", "metadata/rollout_026/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0012.png", "metadata/rollout_026/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0013.png", "metadata/rollout_026/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0014.png", "metadata/rollout_026/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0015.png", "metadata/rollout_026/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0016.png", "metadata/rollout_026/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0017.png", "metadata/rollout_026/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0018.png", "metadata/rollout_026/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0019.png", "metadata/rollout_026/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_026/agentview/0020.png", "metadata/rollout_026/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_026/agentview/0021.png", "metadata/rollout_026/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_026/agentview/0022.png", "metadata/rollout_026/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_026/agentview/0023.png", "metadata/rollout_026/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_026/agentview/0024.png", "metadata/rollout_026/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_026/agentview/0025.png", "metadata/rollout_026/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0026.png", "metadata/rollout_026/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0027.png", "metadata/rollout_026/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0028.png", "metadata/rollout_026/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0029.png", "metadata/rollout_026/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0030.png", "metadata/rollout_026/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_026/agentview/0031.png", "metadata/rollout_026/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_026/agentview/0032.png", "metadata/rollout_026/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_026/agentview/0033.png", "metadata/rollout_026/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_026/agentview/0034.png", "metadata/rollout_026/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_026/agentview/0035.png", "metadata/rollout_026/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_026/agentview/0036.png", "metadata/rollout_026/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_026/agentview/0037.png", "metadata/rollout_026/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_026/agentview/0038.png", "metadata/rollout_026/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0039.png", "metadata/rollout_026/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0040.png", "metadata/rollout_026/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0041.png", "metadata/rollout_026/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0042.png", "metadata/rollout_026/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0043.png", "metadata/rollout_026/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0044.png", "metadata/rollout_026/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0045.png", "metadata/rollout_026/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0046.png", "metadata/rollout_026/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0047.png", "metadata/rollout_026/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0048.png", "metadata/rollout_026/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0049.png", "metadata/rollout_026/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0050.png", "metadata/rollout_026/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0051.png", "metadata/rollout_026/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0052.png", "metadata/rollout_026/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_026/agentview/0053.png", "metadata/rollout_026/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_026/agentview/0054.png", "metadata/rollout_026/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_026/agentview/0055.png", "metadata/rollout_026/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_026/agentview/0056.png", "metadata/rollout_026/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_026/agentview/0057.png", "metadata/rollout_026/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_026/agentview/0058.png", "metadata/rollout_026/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0059.png", "metadata/rollout_026/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0060.png", "metadata/rollout_026/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_026/agentview/0061.png", "metadata/rollout_026/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_026/agentview/0062.png", "metadata/rollout_026/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_026/agentview/0063.png", "metadata/rollout_026/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_026/agentview/0064.png", "metadata/rollout_026/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_026/agentview/0064.png", "metadata/rollout_026/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0000.png", "metadata/rollout_027/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0001.png", "metadata/rollout_027/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0002.png", "metadata/rollout_027/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0003.png", "metadata/rollout_027/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0004.png", "metadata/rollout_027/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0005.png", "metadata/rollout_027/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0006.png", "metadata/rollout_027/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0007.png", "metadata/rollout_027/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0008.png", "metadata/rollout_027/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_027/agentview/0009.png", "metadata/rollout_027/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0010.png", "metadata/rollout_027/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0011.png", "metadata/rollout_027/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0012.png", "metadata/rollout_027/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0013.png", "metadata/rollout_027/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0014.png", "metadata/rollout_027/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0015.png", "metadata/rollout_027/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0016.png", "metadata/rollout_027/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0017.png", "metadata/rollout_027/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0018.png", "metadata/rollout_027/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_027/agentview/0019.png", "metadata/rollout_027/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_027/agentview/0020.png", "metadata/rollout_027/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_027/agentview/0021.png", "metadata/rollout_027/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_027/agentview/0022.png", "metadata/rollout_027/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_027/agentview/0023.png", "metadata/rollout_027/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_027/agentview/0024.png", "metadata/rollout_027/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_027/agentview/0025.png", "metadata/rollout_027/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_027/agentview/0026.png", "metadata/rollout_027/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_027/agentview/0027.png", "metadata/rollout_027/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0028.png", "metadata/rollout_027/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0029.png", "metadata/rollout_027/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0030.png", "metadata/rollout_027/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0031.png", "metadata/rollout_027/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0032.png", "metadata/rollout_027/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0033.png", "metadata/rollout_027/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_027/agentview/0034.png", "metadata/rollout_027/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0035.png", "metadata/rollout_027/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0036.png", "metadata/rollout_027/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0037.png", "metadata/rollout_027/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0038.png", "metadata/rollout_027/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0039.png", "metadata/rollout_027/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0040.png", "metadata/rollout_027/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0041.png", "metadata/rollout_027/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0042.png", "metadata/rollout_027/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0043.png", "metadata/rollout_027/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0044.png", "metadata/rollout_027/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0045.png", "metadata/rollout_027/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0046.png", "metadata/rollout_027/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0047.png", "metadata/rollout_027/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0048.png", "metadata/rollout_027/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0049.png", "metadata/rollout_027/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0050.png", "metadata/rollout_027/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0051.png", "metadata/rollout_027/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0052.png", "metadata/rollout_027/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0053.png", "metadata/rollout_027/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0054.png", "metadata/rollout_027/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0055.png", "metadata/rollout_027/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0056.png", "metadata/rollout_027/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0057.png", "metadata/rollout_027/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_027/agentview/0058.png", "metadata/rollout_027/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0059.png", "metadata/rollout_027/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0060.png", "metadata/rollout_027/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0061.png", "metadata/rollout_027/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_027/agentview/0062.png", "metadata/rollout_027/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_027/agentview/0063.png", "metadata/rollout_027/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0064.png", "metadata/rollout_027/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_027/agentview/0065.png", "metadata/rollout_027/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_027/agentview/0065.png", "metadata/rollout_027/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0000.png", "metadata/rollout_028/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0001.png", "metadata/rollout_028/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0002.png", "metadata/rollout_028/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0003.png", "metadata/rollout_028/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0004.png", "metadata/rollout_028/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0005.png", "metadata/rollout_028/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0006.png", "metadata/rollout_028/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0007.png", "metadata/rollout_028/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0008.png", "metadata/rollout_028/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_028/agentview/0009.png", "metadata/rollout_028/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0010.png", "metadata/rollout_028/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0011.png", "metadata/rollout_028/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0012.png", "metadata/rollout_028/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0013.png", "metadata/rollout_028/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0014.png", "metadata/rollout_028/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0015.png", "metadata/rollout_028/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0016.png", "metadata/rollout_028/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0017.png", "metadata/rollout_028/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0018.png", "metadata/rollout_028/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_028/agentview/0019.png", "metadata/rollout_028/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_028/agentview/0020.png", "metadata/rollout_028/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_028/agentview/0021.png", "metadata/rollout_028/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0022.png", "metadata/rollout_028/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0023.png", "metadata/rollout_028/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0024.png", "metadata/rollout_028/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0025.png", "metadata/rollout_028/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0026.png", "metadata/rollout_028/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0027.png", "metadata/rollout_028/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_028/agentview/0028.png", "metadata/rollout_028/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0029.png", "metadata/rollout_028/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0030.png", "metadata/rollout_028/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0031.png", "metadata/rollout_028/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0032.png", "metadata/rollout_028/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0033.png", "metadata/rollout_028/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0034.png", "metadata/rollout_028/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0035.png", "metadata/rollout_028/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0036.png", "metadata/rollout_028/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0037.png", "metadata/rollout_028/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0038.png", "metadata/rollout_028/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0039.png", "metadata/rollout_028/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0040.png", "metadata/rollout_028/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0041.png", "metadata/rollout_028/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0042.png", "metadata/rollout_028/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0043.png", "metadata/rollout_028/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0044.png", "metadata/rollout_028/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0045.png", "metadata/rollout_028/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_028/agentview/0046.png", "metadata/rollout_028/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0047.png", "metadata/rollout_028/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0048.png", "metadata/rollout_028/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0049.png", "metadata/rollout_028/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_028/agentview/0050.png", "metadata/rollout_028/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_028/agentview/0051.png", "metadata/rollout_028/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0052.png", "metadata/rollout_028/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_028/agentview/0053.png", "metadata/rollout_028/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_028/agentview/0053.png", "metadata/rollout_028/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0000.png", "metadata/rollout_029/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0001.png", "metadata/rollout_029/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0002.png", "metadata/rollout_029/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0003.png", "metadata/rollout_029/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0004.png", "metadata/rollout_029/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0005.png", "metadata/rollout_029/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0006.png", "metadata/rollout_029/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0007.png", "metadata/rollout_029/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0008.png", "metadata/rollout_029/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_029/agentview/0009.png", "metadata/rollout_029/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0010.png", "metadata/rollout_029/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0011.png", "metadata/rollout_029/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0012.png", "metadata/rollout_029/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0013.png", "metadata/rollout_029/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0014.png", "metadata/rollout_029/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0015.png", "metadata/rollout_029/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0016.png", "metadata/rollout_029/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0017.png", "metadata/rollout_029/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0018.png", "metadata/rollout_029/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_029/agentview/0019.png", "metadata/rollout_029/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_029/agentview/0020.png", "metadata/rollout_029/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_029/agentview/0021.png", "metadata/rollout_029/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_029/agentview/0022.png", "metadata/rollout_029/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_029/agentview/0023.png", "metadata/rollout_029/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_029/agentview/0024.png", "metadata/rollout_029/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0025.png", "metadata/rollout_029/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0026.png", "metadata/rollout_029/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0027.png", "metadata/rollout_029/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0028.png", "metadata/rollout_029/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0029.png", "metadata/rollout_029/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0030.png", "metadata/rollout_029/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_029/agentview/0031.png", "metadata/rollout_029/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0032.png", "metadata/rollout_029/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0033.png", "metadata/rollout_029/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0034.png", "metadata/rollout_029/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0035.png", "metadata/rollout_029/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0036.png", "metadata/rollout_029/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0037.png", "metadata/rollout_029/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0038.png", "metadata/rollout_029/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0039.png", "metadata/rollout_029/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0040.png", "metadata/rollout_029/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0041.png", "metadata/rollout_029/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0042.png", "metadata/rollout_029/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0043.png", "metadata/rollout_029/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0044.png", "metadata/rollout_029/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0045.png", "metadata/rollout_029/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0046.png", "metadata/rollout_029/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0047.png", "metadata/rollout_029/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0048.png", "metadata/rollout_029/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0049.png", "metadata/rollout_029/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_029/agentview/0050.png", "metadata/rollout_029/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_029/agentview/0051.png", "metadata/rollout_029/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0052.png", "metadata/rollout_029/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0053.png", "metadata/rollout_029/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0054.png", "metadata/rollout_029/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_029/agentview/0055.png", "metadata/rollout_029/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_029/agentview/0056.png", "metadata/rollout_029/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0057.png", "metadata/rollout_029/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_029/agentview/0058.png", "metadata/rollout_029/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the blue block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_029/agentview/0058.png", "metadata/rollout_029/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0000.png", "metadata/rollout_030/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0001.png", "metadata/rollout_030/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0002.png", "metadata/rollout_030/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0003.png", "metadata/rollout_030/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0004.png", "metadata/rollout_030/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0005.png", "metadata/rollout_030/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0006.png", "metadata/rollout_030/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0007.png", "metadata/rollout_030/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0008.png", "metadata/rollout_030/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0009.png", "metadata/rollout_030/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0010.png", "metadata/rollout_030/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0011.png", "metadata/rollout_030/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0012.png", "metadata/rollout_030/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0013.png", "metadata/rollout_030/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0014.png", "metadata/rollout_030/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0015.png", "metadata/rollout_030/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0016.png", "metadata/rollout_030/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0017.png", "metadata/rollout_030/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0018.png", "metadata/rollout_030/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0019.png", "metadata/rollout_030/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0020.png", "metadata/rollout_030/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0021.png", "metadata/rollout_030/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0022.png", "metadata/rollout_030/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0023.png", "metadata/rollout_030/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0024.png", "metadata/rollout_030/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0025.png", "metadata/rollout_030/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0026.png", "metadata/rollout_030/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0027.png", "metadata/rollout_030/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0028.png", "metadata/rollout_030/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_030/agentview/0029.png", "metadata/rollout_030/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_030/agentview/0030.png", "metadata/rollout_030/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_030/agentview/0031.png", "metadata/rollout_030/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_030/agentview/0032.png", "metadata/rollout_030/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0033.png", "metadata/rollout_030/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0034.png", "metadata/rollout_030/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0035.png", "metadata/rollout_030/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0036.png", "metadata/rollout_030/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0037.png", "metadata/rollout_030/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0038.png", "metadata/rollout_030/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0039.png", "metadata/rollout_030/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0040.png", "metadata/rollout_030/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0041.png", "metadata/rollout_030/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0042.png", "metadata/rollout_030/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0043.png", "metadata/rollout_030/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0044.png", "metadata/rollout_030/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0045.png", "metadata/rollout_030/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0046.png", "metadata/rollout_030/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0047.png", "metadata/rollout_030/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0048.png", "metadata/rollout_030/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0049.png", "metadata/rollout_030/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0050.png", "metadata/rollout_030/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0051.png", "metadata/rollout_030/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0052.png", "metadata/rollout_030/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0053.png", "metadata/rollout_030/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0054.png", "metadata/rollout_030/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0055.png", "metadata/rollout_030/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_030/agentview/0056.png", "metadata/rollout_030/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_030/agentview/0057.png", "metadata/rollout_030/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0058.png", "metadata/rollout_030/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_030/agentview/0059.png", "metadata/rollout_030/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_030/agentview/0060.png", "metadata/rollout_030/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0061.png", "metadata/rollout_030/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_030/agentview/0062.png", "metadata/rollout_030/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_030/agentview/0062.png", "metadata/rollout_030/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0000.png", "metadata/rollout_031/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0001.png", "metadata/rollout_031/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0002.png", "metadata/rollout_031/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0003.png", "metadata/rollout_031/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0004.png", "metadata/rollout_031/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0005.png", "metadata/rollout_031/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0006.png", "metadata/rollout_031/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0007.png", "metadata/rollout_031/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0008.png", "metadata/rollout_031/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0009.png", "metadata/rollout_031/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0010.png", "metadata/rollout_031/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0011.png", "metadata/rollout_031/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0012.png", "metadata/rollout_031/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0013.png", "metadata/rollout_031/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0014.png", "metadata/rollout_031/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0015.png", "metadata/rollout_031/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0016.png", "metadata/rollout_031/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0017.png", "metadata/rollout_031/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0018.png", "metadata/rollout_031/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0019.png", "metadata/rollout_031/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0020.png", "metadata/rollout_031/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0021.png", "metadata/rollout_031/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0022.png", "metadata/rollout_031/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0023.png", "metadata/rollout_031/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_031/agentview/0024.png", "metadata/rollout_031/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0025.png", "metadata/rollout_031/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0026.png", "metadata/rollout_031/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0027.png", "metadata/rollout_031/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0028.png", "metadata/rollout_031/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0029.png", "metadata/rollout_031/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0030.png", "metadata/rollout_031/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0031.png", "metadata/rollout_031/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0032.png", "metadata/rollout_031/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0033.png", "metadata/rollout_031/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0034.png", "metadata/rollout_031/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0035.png", "metadata/rollout_031/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0036.png", "metadata/rollout_031/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0037.png", "metadata/rollout_031/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0038.png", "metadata/rollout_031/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0039.png", "metadata/rollout_031/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_031/agentview/0040.png", "metadata/rollout_031/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0041.png", "metadata/rollout_031/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0042.png", "metadata/rollout_031/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0043.png", "metadata/rollout_031/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0044.png", "metadata/rollout_031/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0045.png", "metadata/rollout_031/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0046.png", "metadata/rollout_031/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0047.png", "metadata/rollout_031/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0048.png", "metadata/rollout_031/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0049.png", "metadata/rollout_031/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0050.png", "metadata/rollout_031/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0051.png", "metadata/rollout_031/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0052.png", "metadata/rollout_031/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0053.png", "metadata/rollout_031/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0054.png", "metadata/rollout_031/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0055.png", "metadata/rollout_031/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0056.png", "metadata/rollout_031/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_031/agentview/0057.png", "metadata/rollout_031/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0058.png", "metadata/rollout_031/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_031/agentview/0059.png", "metadata/rollout_031/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_031/agentview/0060.png", "metadata/rollout_031/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0061.png", "metadata/rollout_031/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_031/agentview/0062.png", "metadata/rollout_031/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_031/agentview/0062.png", "metadata/rollout_031/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0000.png", "metadata/rollout_032/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0001.png", "metadata/rollout_032/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0002.png", "metadata/rollout_032/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0003.png", "metadata/rollout_032/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0004.png", "metadata/rollout_032/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0005.png", "metadata/rollout_032/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0006.png", "metadata/rollout_032/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0007.png", "metadata/rollout_032/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0008.png", "metadata/rollout_032/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0009.png", "metadata/rollout_032/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_032/agentview/0010.png", "metadata/rollout_032/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_032/agentview/0011.png", "metadata/rollout_032/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_032/agentview/0012.png", "metadata/rollout_032/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_032/agentview/0013.png", "metadata/rollout_032/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_032/agentview/0014.png", "metadata/rollout_032/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_032/agentview/0015.png", "metadata/rollout_032/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_032/agentview/0016.png", "metadata/rollout_032/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_032/agentview/0017.png", "metadata/rollout_032/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0018.png", "metadata/rollout_032/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0019.png", "metadata/rollout_032/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0020.png", "metadata/rollout_032/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0021.png", "metadata/rollout_032/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0022.png", "metadata/rollout_032/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_032/agentview/0023.png", "metadata/rollout_032/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0024.png", "metadata/rollout_032/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0025.png", "metadata/rollout_032/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0026.png", "metadata/rollout_032/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0027.png", "metadata/rollout_032/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0028.png", "metadata/rollout_032/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0029.png", "metadata/rollout_032/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0030.png", "metadata/rollout_032/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0031.png", "metadata/rollout_032/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0032.png", "metadata/rollout_032/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0033.png", "metadata/rollout_032/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0034.png", "metadata/rollout_032/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0035.png", "metadata/rollout_032/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0036.png", "metadata/rollout_032/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0037.png", "metadata/rollout_032/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0038.png", "metadata/rollout_032/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0039.png", "metadata/rollout_032/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0040.png", "metadata/rollout_032/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0041.png", "metadata/rollout_032/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0042.png", "metadata/rollout_032/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0043.png", "metadata/rollout_032/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0044.png", "metadata/rollout_032/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0045.png", "metadata/rollout_032/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0046.png", "metadata/rollout_032/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0047.png", "metadata/rollout_032/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0048.png", "metadata/rollout_032/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_032/agentview/0049.png", "metadata/rollout_032/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_032/agentview/0050.png", "metadata/rollout_032/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0051.png", "metadata/rollout_032/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0052.png", "metadata/rollout_032/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_032/agentview/0053.png", "metadata/rollout_032/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_032/agentview/0054.png", "metadata/rollout_032/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0055.png", "metadata/rollout_032/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_032/agentview/0056.png", "metadata/rollout_032/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_032/agentview/0056.png", "metadata/rollout_032/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0000.png", "metadata/rollout_033/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0001.png", "metadata/rollout_033/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0002.png", "metadata/rollout_033/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0003.png", "metadata/rollout_033/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0004.png", "metadata/rollout_033/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0005.png", "metadata/rollout_033/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0006.png", "metadata/rollout_033/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0007.png", "metadata/rollout_033/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0008.png", "metadata/rollout_033/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0009.png", "metadata/rollout_033/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0010.png", "metadata/rollout_033/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0011.png", "metadata/rollout_033/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0012.png", "metadata/rollout_033/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0013.png", "metadata/rollout_033/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0014.png", "metadata/rollout_033/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0015.png", "metadata/rollout_033/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0016.png", "metadata/rollout_033/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0017.png", "metadata/rollout_033/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0018.png", "metadata/rollout_033/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0019.png", "metadata/rollout_033/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0020.png", "metadata/rollout_033/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0021.png", "metadata/rollout_033/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0022.png", "metadata/rollout_033/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0023.png", "metadata/rollout_033/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0024.png", "metadata/rollout_033/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0025.png", "metadata/rollout_033/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0026.png", "metadata/rollout_033/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0027.png", "metadata/rollout_033/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0028.png", "metadata/rollout_033/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0029.png", "metadata/rollout_033/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0030.png", "metadata/rollout_033/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_033/agentview/0031.png", "metadata/rollout_033/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_033/agentview/0032.png", "metadata/rollout_033/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_033/agentview/0033.png", "metadata/rollout_033/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_033/agentview/0034.png", "metadata/rollout_033/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_033/agentview/0035.png", "metadata/rollout_033/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_033/agentview/0036.png", "metadata/rollout_033/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_033/agentview/0037.png", "metadata/rollout_033/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0038.png", "metadata/rollout_033/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_033/agentview/0039.png", "metadata/rollout_033/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0040.png", "metadata/rollout_033/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0041.png", "metadata/rollout_033/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0042.png", "metadata/rollout_033/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0043.png", "metadata/rollout_033/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0044.png", "metadata/rollout_033/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0045.png", "metadata/rollout_033/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0046.png", "metadata/rollout_033/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0047.png", "metadata/rollout_033/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0048.png", "metadata/rollout_033/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0049.png", "metadata/rollout_033/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_033/agentview/0050.png", "metadata/rollout_033/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0051.png", "metadata/rollout_033/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0052.png", "metadata/rollout_033/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0053.png", "metadata/rollout_033/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0054.png", "metadata/rollout_033/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0055.png", "metadata/rollout_033/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0056.png", "metadata/rollout_033/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0057.png", "metadata/rollout_033/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0058.png", "metadata/rollout_033/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0059.png", "metadata/rollout_033/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0060.png", "metadata/rollout_033/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0061.png", "metadata/rollout_033/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_033/agentview/0062.png", "metadata/rollout_033/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0063.png", "metadata/rollout_033/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_033/agentview/0064.png", "metadata/rollout_033/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_033/agentview/0065.png", "metadata/rollout_033/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0066.png", "metadata/rollout_033/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_033/agentview/0067.png", "metadata/rollout_033/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_033/agentview/0067.png", "metadata/rollout_033/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0000.png", "metadata/rollout_034/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0001.png", "metadata/rollout_034/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0002.png", "metadata/rollout_034/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0003.png", "metadata/rollout_034/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0004.png", "metadata/rollout_034/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0005.png", "metadata/rollout_034/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0006.png", "metadata/rollout_034/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0007.png", "metadata/rollout_034/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0008.png", "metadata/rollout_034/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0009.png", "metadata/rollout_034/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0010.png", "metadata/rollout_034/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0011.png", "metadata/rollout_034/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0012.png", "metadata/rollout_034/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0013.png", "metadata/rollout_034/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0014.png", "metadata/rollout_034/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0015.png", "metadata/rollout_034/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0016.png", "metadata/rollout_034/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0017.png", "metadata/rollout_034/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0018.png", "metadata/rollout_034/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0019.png", "metadata/rollout_034/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0020.png", "metadata/rollout_034/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0021.png", "metadata/rollout_034/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0022.png", "metadata/rollout_034/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0023.png", "metadata/rollout_034/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0024.png", "metadata/rollout_034/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_034/agentview/0025.png", "metadata/rollout_034/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0026.png", "metadata/rollout_034/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0027.png", "metadata/rollout_034/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0028.png", "metadata/rollout_034/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0029.png", "metadata/rollout_034/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0030.png", "metadata/rollout_034/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0031.png", "metadata/rollout_034/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0032.png", "metadata/rollout_034/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0033.png", "metadata/rollout_034/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0034.png", "metadata/rollout_034/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0035.png", "metadata/rollout_034/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0036.png", "metadata/rollout_034/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0037.png", "metadata/rollout_034/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0038.png", "metadata/rollout_034/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0039.png", "metadata/rollout_034/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0040.png", "metadata/rollout_034/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0041.png", "metadata/rollout_034/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0042.png", "metadata/rollout_034/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0043.png", "metadata/rollout_034/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0044.png", "metadata/rollout_034/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0045.png", "metadata/rollout_034/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0046.png", "metadata/rollout_034/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0047.png", "metadata/rollout_034/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0048.png", "metadata/rollout_034/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0049.png", "metadata/rollout_034/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_034/agentview/0050.png", "metadata/rollout_034/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0051.png", "metadata/rollout_034/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0052.png", "metadata/rollout_034/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_034/agentview/0053.png", "metadata/rollout_034/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0054.png", "metadata/rollout_034/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0055.png", "metadata/rollout_034/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_034/agentview/0056.png", "metadata/rollout_034/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0057.png", "metadata/rollout_034/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_034/agentview/0058.png", "metadata/rollout_034/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0059.png", "metadata/rollout_034/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_034/agentview/0060.png", "metadata/rollout_034/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_034/agentview/0060.png", "metadata/rollout_034/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0000.png", "metadata/rollout_035/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0001.png", "metadata/rollout_035/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0002.png", "metadata/rollout_035/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0003.png", "metadata/rollout_035/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0004.png", "metadata/rollout_035/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0005.png", "metadata/rollout_035/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0006.png", "metadata/rollout_035/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0007.png", "metadata/rollout_035/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0008.png", "metadata/rollout_035/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0009.png", "metadata/rollout_035/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0010.png", "metadata/rollout_035/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0011.png", "metadata/rollout_035/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0012.png", "metadata/rollout_035/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0013.png", "metadata/rollout_035/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0014.png", "metadata/rollout_035/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0015.png", "metadata/rollout_035/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0016.png", "metadata/rollout_035/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0017.png", "metadata/rollout_035/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0018.png", "metadata/rollout_035/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0019.png", "metadata/rollout_035/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0020.png", "metadata/rollout_035/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0021.png", "metadata/rollout_035/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0022.png", "metadata/rollout_035/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0023.png", "metadata/rollout_035/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0024.png", "metadata/rollout_035/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0025.png", "metadata/rollout_035/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_035/agentview/0026.png", "metadata/rollout_035/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0027.png", "metadata/rollout_035/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0028.png", "metadata/rollout_035/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0029.png", "metadata/rollout_035/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0030.png", "metadata/rollout_035/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0031.png", "metadata/rollout_035/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0032.png", "metadata/rollout_035/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0033.png", "metadata/rollout_035/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0034.png", "metadata/rollout_035/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0035.png", "metadata/rollout_035/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0036.png", "metadata/rollout_035/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0037.png", "metadata/rollout_035/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0038.png", "metadata/rollout_035/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0039.png", "metadata/rollout_035/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0040.png", "metadata/rollout_035/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0041.png", "metadata/rollout_035/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0042.png", "metadata/rollout_035/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0043.png", "metadata/rollout_035/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0044.png", "metadata/rollout_035/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0045.png", "metadata/rollout_035/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0046.png", "metadata/rollout_035/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0047.png", "metadata/rollout_035/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0048.png", "metadata/rollout_035/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0049.png", "metadata/rollout_035/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0050.png", "metadata/rollout_035/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0051.png", "metadata/rollout_035/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_035/agentview/0052.png", "metadata/rollout_035/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0053.png", "metadata/rollout_035/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0054.png", "metadata/rollout_035/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0055.png", "metadata/rollout_035/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0056.png", "metadata/rollout_035/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0057.png", "metadata/rollout_035/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0058.png", "metadata/rollout_035/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_035/agentview/0059.png", "metadata/rollout_035/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0060.png", "metadata/rollout_035/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_035/agentview/0061.png", "metadata/rollout_035/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_035/agentview/0062.png", "metadata/rollout_035/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0063.png", "metadata/rollout_035/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_035/agentview/0064.png", "metadata/rollout_035/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_035/agentview/0064.png", "metadata/rollout_035/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0000.png", "metadata/rollout_036/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0001.png", "metadata/rollout_036/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0002.png", "metadata/rollout_036/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0003.png", "metadata/rollout_036/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0004.png", "metadata/rollout_036/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0005.png", "metadata/rollout_036/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0006.png", "metadata/rollout_036/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0007.png", "metadata/rollout_036/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0008.png", "metadata/rollout_036/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0009.png", "metadata/rollout_036/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0010.png", "metadata/rollout_036/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0011.png", "metadata/rollout_036/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0012.png", "metadata/rollout_036/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0013.png", "metadata/rollout_036/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0014.png", "metadata/rollout_036/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0015.png", "metadata/rollout_036/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0016.png", "metadata/rollout_036/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0017.png", "metadata/rollout_036/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0018.png", "metadata/rollout_036/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0019.png", "metadata/rollout_036/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0020.png", "metadata/rollout_036/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0021.png", "metadata/rollout_036/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0022.png", "metadata/rollout_036/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0023.png", "metadata/rollout_036/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0024.png", "metadata/rollout_036/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0025.png", "metadata/rollout_036/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0026.png", "metadata/rollout_036/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0027.png", "metadata/rollout_036/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0028.png", "metadata/rollout_036/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_036/agentview/0029.png", "metadata/rollout_036/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0030.png", "metadata/rollout_036/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0031.png", "metadata/rollout_036/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0032.png", "metadata/rollout_036/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0033.png", "metadata/rollout_036/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0034.png", "metadata/rollout_036/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0035.png", "metadata/rollout_036/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0036.png", "metadata/rollout_036/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0037.png", "metadata/rollout_036/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0038.png", "metadata/rollout_036/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0039.png", "metadata/rollout_036/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0040.png", "metadata/rollout_036/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0041.png", "metadata/rollout_036/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0042.png", "metadata/rollout_036/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0043.png", "metadata/rollout_036/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0044.png", "metadata/rollout_036/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0045.png", "metadata/rollout_036/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0046.png", "metadata/rollout_036/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0047.png", "metadata/rollout_036/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0048.png", "metadata/rollout_036/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0049.png", "metadata/rollout_036/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0050.png", "metadata/rollout_036/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_036/agentview/0051.png", "metadata/rollout_036/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0052.png", "metadata/rollout_036/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0053.png", "metadata/rollout_036/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0054.png", "metadata/rollout_036/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_036/agentview/0055.png", "metadata/rollout_036/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0056.png", "metadata/rollout_036/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0057.png", "metadata/rollout_036/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_036/agentview/0058.png", "metadata/rollout_036/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0059.png", "metadata/rollout_036/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_036/agentview/0060.png", "metadata/rollout_036/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0061.png", "metadata/rollout_036/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_036/agentview/0062.png", "metadata/rollout_036/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_036/agentview/0062.png", "metadata/rollout_036/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0000.png", "metadata/rollout_037/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0001.png", "metadata/rollout_037/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0002.png", "metadata/rollout_037/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0003.png", "metadata/rollout_037/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0004.png", "metadata/rollout_037/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0005.png", "metadata/rollout_037/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0006.png", "metadata/rollout_037/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0007.png", "metadata/rollout_037/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0008.png", "metadata/rollout_037/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0009.png", "metadata/rollout_037/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0010.png", "metadata/rollout_037/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0011.png", "metadata/rollout_037/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0012.png", "metadata/rollout_037/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0013.png", "metadata/rollout_037/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0014.png", "metadata/rollout_037/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0015.png", "metadata/rollout_037/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0016.png", "metadata/rollout_037/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0017.png", "metadata/rollout_037/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0018.png", "metadata/rollout_037/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0019.png", "metadata/rollout_037/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_037/agentview/0020.png", "metadata/rollout_037/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_037/agentview/0021.png", "metadata/rollout_037/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_037/agentview/0022.png", "metadata/rollout_037/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0023.png", "metadata/rollout_037/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0024.png", "metadata/rollout_037/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0025.png", "metadata/rollout_037/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0026.png", "metadata/rollout_037/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0027.png", "metadata/rollout_037/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0028.png", "metadata/rollout_037/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_037/agentview/0029.png", "metadata/rollout_037/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0030.png", "metadata/rollout_037/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0031.png", "metadata/rollout_037/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0032.png", "metadata/rollout_037/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0033.png", "metadata/rollout_037/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0034.png", "metadata/rollout_037/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0035.png", "metadata/rollout_037/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0036.png", "metadata/rollout_037/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0037.png", "metadata/rollout_037/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0038.png", "metadata/rollout_037/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0039.png", "metadata/rollout_037/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0040.png", "metadata/rollout_037/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0041.png", "metadata/rollout_037/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0042.png", "metadata/rollout_037/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0043.png", "metadata/rollout_037/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0044.png", "metadata/rollout_037/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0045.png", "metadata/rollout_037/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0046.png", "metadata/rollout_037/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0047.png", "metadata/rollout_037/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0048.png", "metadata/rollout_037/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0049.png", "metadata/rollout_037/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0050.png", "metadata/rollout_037/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0051.png", "metadata/rollout_037/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_037/agentview/0052.png", "metadata/rollout_037/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0053.png", "metadata/rollout_037/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0054.png", "metadata/rollout_037/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0055.png", "metadata/rollout_037/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_037/agentview/0056.png", "metadata/rollout_037/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0057.png", "metadata/rollout_037/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_037/agentview/0058.png", "metadata/rollout_037/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_037/agentview/0059.png", "metadata/rollout_037/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0060.png", "metadata/rollout_037/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_037/agentview/0061.png", "metadata/rollout_037/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_037/agentview/0061.png", "metadata/rollout_037/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0000.png", "metadata/rollout_038/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0001.png", "metadata/rollout_038/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0002.png", "metadata/rollout_038/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0003.png", "metadata/rollout_038/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0004.png", "metadata/rollout_038/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0005.png", "metadata/rollout_038/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0006.png", "metadata/rollout_038/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0007.png", "metadata/rollout_038/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0008.png", "metadata/rollout_038/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0009.png", "metadata/rollout_038/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0010.png", "metadata/rollout_038/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0011.png", "metadata/rollout_038/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0012.png", "metadata/rollout_038/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0013.png", "metadata/rollout_038/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0014.png", "metadata/rollout_038/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0015.png", "metadata/rollout_038/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0016.png", "metadata/rollout_038/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0017.png", "metadata/rollout_038/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_038/agentview/0018.png", "metadata/rollout_038/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_038/agentview/0019.png", "metadata/rollout_038/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_038/agentview/0020.png", "metadata/rollout_038/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_038/agentview/0021.png", "metadata/rollout_038/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0022.png", "metadata/rollout_038/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0023.png", "metadata/rollout_038/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0024.png", "metadata/rollout_038/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0025.png", "metadata/rollout_038/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0026.png", "metadata/rollout_038/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_038/agentview/0027.png", "metadata/rollout_038/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0028.png", "metadata/rollout_038/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0029.png", "metadata/rollout_038/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0030.png", "metadata/rollout_038/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0031.png", "metadata/rollout_038/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0032.png", "metadata/rollout_038/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0033.png", "metadata/rollout_038/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0034.png", "metadata/rollout_038/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0035.png", "metadata/rollout_038/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0036.png", "metadata/rollout_038/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0037.png", "metadata/rollout_038/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0038.png", "metadata/rollout_038/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0039.png", "metadata/rollout_038/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0040.png", "metadata/rollout_038/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0041.png", "metadata/rollout_038/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0042.png", "metadata/rollout_038/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0043.png", "metadata/rollout_038/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0044.png", "metadata/rollout_038/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0045.png", "metadata/rollout_038/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0046.png", "metadata/rollout_038/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0047.png", "metadata/rollout_038/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0048.png", "metadata/rollout_038/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0049.png", "metadata/rollout_038/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0050.png", "metadata/rollout_038/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0051.png", "metadata/rollout_038/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0052.png", "metadata/rollout_038/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0053.png", "metadata/rollout_038/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0054.png", "metadata/rollout_038/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0055.png", "metadata/rollout_038/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_038/agentview/0056.png", "metadata/rollout_038/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_038/agentview/0057.png", "metadata/rollout_038/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0058.png", "metadata/rollout_038/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_038/agentview/0059.png", "metadata/rollout_038/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_038/agentview/0060.png", "metadata/rollout_038/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0061.png", "metadata/rollout_038/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_038/agentview/0062.png", "metadata/rollout_038/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_038/agentview/0062.png", "metadata/rollout_038/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0000.png", "metadata/rollout_039/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0001.png", "metadata/rollout_039/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0002.png", "metadata/rollout_039/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0003.png", "metadata/rollout_039/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0004.png", "metadata/rollout_039/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0005.png", "metadata/rollout_039/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0006.png", "metadata/rollout_039/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0007.png", "metadata/rollout_039/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0008.png", "metadata/rollout_039/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0009.png", "metadata/rollout_039/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0010.png", "metadata/rollout_039/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0011.png", "metadata/rollout_039/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0012.png", "metadata/rollout_039/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0013.png", "metadata/rollout_039/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0014.png", "metadata/rollout_039/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0015.png", "metadata/rollout_039/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0016.png", "metadata/rollout_039/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0017.png", "metadata/rollout_039/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_039/agentview/0018.png", "metadata/rollout_039/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_039/agentview/0019.png", "metadata/rollout_039/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_039/agentview/0020.png", "metadata/rollout_039/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_039/agentview/0021.png", "metadata/rollout_039/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0022.png", "metadata/rollout_039/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0023.png", "metadata/rollout_039/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0024.png", "metadata/rollout_039/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0025.png", "metadata/rollout_039/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0026.png", "metadata/rollout_039/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_039/agentview/0027.png", "metadata/rollout_039/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0028.png", "metadata/rollout_039/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0029.png", "metadata/rollout_039/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0030.png", "metadata/rollout_039/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0031.png", "metadata/rollout_039/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0032.png", "metadata/rollout_039/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0033.png", "metadata/rollout_039/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0034.png", "metadata/rollout_039/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0035.png", "metadata/rollout_039/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0036.png", "metadata/rollout_039/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0037.png", "metadata/rollout_039/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0038.png", "metadata/rollout_039/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0039.png", "metadata/rollout_039/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0040.png", "metadata/rollout_039/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0041.png", "metadata/rollout_039/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0042.png", "metadata/rollout_039/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0043.png", "metadata/rollout_039/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0044.png", "metadata/rollout_039/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0045.png", "metadata/rollout_039/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0046.png", "metadata/rollout_039/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0047.png", "metadata/rollout_039/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0048.png", "metadata/rollout_039/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0049.png", "metadata/rollout_039/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0050.png", "metadata/rollout_039/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0051.png", "metadata/rollout_039/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0052.png", "metadata/rollout_039/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0053.png", "metadata/rollout_039/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0054.png", "metadata/rollout_039/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0055.png", "metadata/rollout_039/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_039/agentview/0056.png", "metadata/rollout_039/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0057.png", "metadata/rollout_039/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0058.png", "metadata/rollout_039/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_039/agentview/0059.png", "metadata/rollout_039/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0060.png", "metadata/rollout_039/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_039/agentview/0061.png", "metadata/rollout_039/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0062.png", "metadata/rollout_039/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_039/agentview/0063.png", "metadata/rollout_039/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0064.png", "metadata/rollout_039/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_039/agentview/0065.png", "metadata/rollout_039/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the canned food in the grey bin\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_039/agentview/0065.png", "metadata/rollout_039/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0000.png", "metadata/rollout_040/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0001.png", "metadata/rollout_040/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0002.png", "metadata/rollout_040/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0003.png", "metadata/rollout_040/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0004.png", "metadata/rollout_040/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0005.png", "metadata/rollout_040/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0006.png", "metadata/rollout_040/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0007.png", "metadata/rollout_040/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0008.png", "metadata/rollout_040/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_040/agentview/0009.png", "metadata/rollout_040/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_040/agentview/0010.png", "metadata/rollout_040/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_040/agentview/0011.png", "metadata/rollout_040/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_040/agentview/0012.png", "metadata/rollout_040/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_040/agentview/0013.png", "metadata/rollout_040/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_040/agentview/0014.png", "metadata/rollout_040/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_040/agentview/0015.png", "metadata/rollout_040/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0016.png", "metadata/rollout_040/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0017.png", "metadata/rollout_040/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0018.png", "metadata/rollout_040/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0019.png", "metadata/rollout_040/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0020.png", "metadata/rollout_040/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0021.png", "metadata/rollout_040/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_040/agentview/0022.png", "metadata/rollout_040/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0023.png", "metadata/rollout_040/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0024.png", "metadata/rollout_040/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0025.png", "metadata/rollout_040/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0026.png", "metadata/rollout_040/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0027.png", "metadata/rollout_040/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0028.png", "metadata/rollout_040/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0029.png", "metadata/rollout_040/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0030.png", "metadata/rollout_040/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_040/agentview/0031.png", "metadata/rollout_040/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_040/agentview/0032.png", "metadata/rollout_040/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_040/agentview/0033.png", "metadata/rollout_040/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_040/agentview/0034.png", "metadata/rollout_040/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_040/agentview/0035.png", "metadata/rollout_040/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_040/agentview/0036.png", "metadata/rollout_040/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_040/agentview/0037.png", "metadata/rollout_040/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_040/agentview/0038.png", "metadata/rollout_040/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_040/agentview/0039.png", "metadata/rollout_040/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0040.png", "metadata/rollout_040/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0041.png", "metadata/rollout_040/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0042.png", "metadata/rollout_040/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_040/agentview/0043.png", "metadata/rollout_040/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_040/agentview/0044.png", "metadata/rollout_040/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0045.png", "metadata/rollout_040/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_040/agentview/0046.png", "metadata/rollout_040/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_040/agentview/0046.png", "metadata/rollout_040/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0000.png", "metadata/rollout_041/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0001.png", "metadata/rollout_041/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0002.png", "metadata/rollout_041/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0003.png", "metadata/rollout_041/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0004.png", "metadata/rollout_041/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0005.png", "metadata/rollout_041/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0006.png", "metadata/rollout_041/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0007.png", "metadata/rollout_041/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0008.png", "metadata/rollout_041/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_041/agentview/0009.png", "metadata/rollout_041/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_041/agentview/0010.png", "metadata/rollout_041/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_041/agentview/0011.png", "metadata/rollout_041/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_041/agentview/0012.png", "metadata/rollout_041/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_041/agentview/0013.png", "metadata/rollout_041/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_041/agentview/0014.png", "metadata/rollout_041/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_041/agentview/0015.png", "metadata/rollout_041/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_041/agentview/0016.png", "metadata/rollout_041/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_041/agentview/0017.png", "metadata/rollout_041/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0018.png", "metadata/rollout_041/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0019.png", "metadata/rollout_041/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0020.png", "metadata/rollout_041/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0021.png", "metadata/rollout_041/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0022.png", "metadata/rollout_041/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0023.png", "metadata/rollout_041/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_041/agentview/0024.png", "metadata/rollout_041/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0025.png", "metadata/rollout_041/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0026.png", "metadata/rollout_041/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0027.png", "metadata/rollout_041/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0028.png", "metadata/rollout_041/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0029.png", "metadata/rollout_041/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0030.png", "metadata/rollout_041/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0031.png", "metadata/rollout_041/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0032.png", "metadata/rollout_041/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0033.png", "metadata/rollout_041/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0034.png", "metadata/rollout_041/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0035.png", "metadata/rollout_041/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0036.png", "metadata/rollout_041/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0037.png", "metadata/rollout_041/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0038.png", "metadata/rollout_041/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0039.png", "metadata/rollout_041/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0040.png", "metadata/rollout_041/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0041.png", "metadata/rollout_041/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0042.png", "metadata/rollout_041/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_041/agentview/0043.png", "metadata/rollout_041/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_041/agentview/0044.png", "metadata/rollout_041/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_041/agentview/0045.png", "metadata/rollout_041/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_041/agentview/0046.png", "metadata/rollout_041/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0047.png", "metadata/rollout_041/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0048.png", "metadata/rollout_041/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0049.png", "metadata/rollout_041/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_041/agentview/0050.png", "metadata/rollout_041/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_041/agentview/0051.png", "metadata/rollout_041/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0052.png", "metadata/rollout_041/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_041/agentview/0053.png", "metadata/rollout_041/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_041/agentview/0053.png", "metadata/rollout_041/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0000.png", "metadata/rollout_042/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0001.png", "metadata/rollout_042/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0002.png", "metadata/rollout_042/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0003.png", "metadata/rollout_042/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0004.png", "metadata/rollout_042/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0005.png", "metadata/rollout_042/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0006.png", "metadata/rollout_042/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0007.png", "metadata/rollout_042/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0008.png", "metadata/rollout_042/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_042/agentview/0009.png", "metadata/rollout_042/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_042/agentview/0010.png", "metadata/rollout_042/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_042/agentview/0011.png", "metadata/rollout_042/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_042/agentview/0012.png", "metadata/rollout_042/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_042/agentview/0013.png", "metadata/rollout_042/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_042/agentview/0014.png", "metadata/rollout_042/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_042/agentview/0015.png", "metadata/rollout_042/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_042/agentview/0016.png", "metadata/rollout_042/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_042/agentview/0017.png", "metadata/rollout_042/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_042/agentview/0018.png", "metadata/rollout_042/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_042/agentview/0019.png", "metadata/rollout_042/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_042/agentview/0020.png", "metadata/rollout_042/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_042/agentview/0021.png", "metadata/rollout_042/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_042/agentview/0022.png", "metadata/rollout_042/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_042/agentview/0023.png", "metadata/rollout_042/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_042/agentview/0024.png", "metadata/rollout_042/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_042/agentview/0025.png", "metadata/rollout_042/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0026.png", "metadata/rollout_042/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0027.png", "metadata/rollout_042/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0028.png", "metadata/rollout_042/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0029.png", "metadata/rollout_042/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0030.png", "metadata/rollout_042/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0031.png", "metadata/rollout_042/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_042/agentview/0032.png", "metadata/rollout_042/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0033.png", "metadata/rollout_042/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0034.png", "metadata/rollout_042/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0035.png", "metadata/rollout_042/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0036.png", "metadata/rollout_042/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0037.png", "metadata/rollout_042/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0038.png", "metadata/rollout_042/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0039.png", "metadata/rollout_042/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0040.png", "metadata/rollout_042/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_042/agentview/0041.png", "metadata/rollout_042/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_042/agentview/0042.png", "metadata/rollout_042/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_042/agentview/0043.png", "metadata/rollout_042/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0044.png", "metadata/rollout_042/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0045.png", "metadata/rollout_042/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0046.png", "metadata/rollout_042/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0047.png", "metadata/rollout_042/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0048.png", "metadata/rollout_042/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0049.png", "metadata/rollout_042/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0050.png", "metadata/rollout_042/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0051.png", "metadata/rollout_042/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0052.png", "metadata/rollout_042/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0053.png", "metadata/rollout_042/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0054.png", "metadata/rollout_042/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0055.png", "metadata/rollout_042/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_042/agentview/0056.png", "metadata/rollout_042/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0057.png", "metadata/rollout_042/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0058.png", "metadata/rollout_042/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0059.png", "metadata/rollout_042/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_042/agentview/0060.png", "metadata/rollout_042/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_042/agentview/0061.png", "metadata/rollout_042/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0062.png", "metadata/rollout_042/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_042/agentview/0063.png", "metadata/rollout_042/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_042/agentview/0063.png", "metadata/rollout_042/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0000.png", "metadata/rollout_043/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0001.png", "metadata/rollout_043/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0002.png", "metadata/rollout_043/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0003.png", "metadata/rollout_043/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0004.png", "metadata/rollout_043/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0005.png", "metadata/rollout_043/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0006.png", "metadata/rollout_043/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0007.png", "metadata/rollout_043/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0008.png", "metadata/rollout_043/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0009.png", "metadata/rollout_043/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0010.png", "metadata/rollout_043/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0011.png", "metadata/rollout_043/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0012.png", "metadata/rollout_043/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0013.png", "metadata/rollout_043/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0014.png", "metadata/rollout_043/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0015.png", "metadata/rollout_043/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0016.png", "metadata/rollout_043/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0017.png", "metadata/rollout_043/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_043/agentview/0018.png", "metadata/rollout_043/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_043/agentview/0019.png", "metadata/rollout_043/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_043/agentview/0020.png", "metadata/rollout_043/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_043/agentview/0021.png", "metadata/rollout_043/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_043/agentview/0022.png", "metadata/rollout_043/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_043/agentview/0023.png", "metadata/rollout_043/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_043/agentview/0024.png", "metadata/rollout_043/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0025.png", "metadata/rollout_043/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0026.png", "metadata/rollout_043/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0027.png", "metadata/rollout_043/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0028.png", "metadata/rollout_043/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0029.png", "metadata/rollout_043/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0030.png", "metadata/rollout_043/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_043/agentview/0031.png", "metadata/rollout_043/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0032.png", "metadata/rollout_043/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0033.png", "metadata/rollout_043/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0034.png", "metadata/rollout_043/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0035.png", "metadata/rollout_043/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0036.png", "metadata/rollout_043/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0037.png", "metadata/rollout_043/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0038.png", "metadata/rollout_043/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0039.png", "metadata/rollout_043/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0040.png", "metadata/rollout_043/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0041.png", "metadata/rollout_043/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0042.png", "metadata/rollout_043/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0043.png", "metadata/rollout_043/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0044.png", "metadata/rollout_043/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0045.png", "metadata/rollout_043/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0046.png", "metadata/rollout_043/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0047.png", "metadata/rollout_043/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0048.png", "metadata/rollout_043/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0049.png", "metadata/rollout_043/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0050.png", "metadata/rollout_043/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0051.png", "metadata/rollout_043/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0052.png", "metadata/rollout_043/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_043/agentview/0053.png", "metadata/rollout_043/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_043/agentview/0054.png", "metadata/rollout_043/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0055.png", "metadata/rollout_043/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0056.png", "metadata/rollout_043/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0057.png", "metadata/rollout_043/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_043/agentview/0058.png", "metadata/rollout_043/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_043/agentview/0059.png", "metadata/rollout_043/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0060.png", "metadata/rollout_043/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_043/agentview/0061.png", "metadata/rollout_043/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_043/agentview/0061.png", "metadata/rollout_043/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0000.png", "metadata/rollout_044/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0001.png", "metadata/rollout_044/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0002.png", "metadata/rollout_044/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0003.png", "metadata/rollout_044/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0004.png", "metadata/rollout_044/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0005.png", "metadata/rollout_044/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0006.png", "metadata/rollout_044/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0007.png", "metadata/rollout_044/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0008.png", "metadata/rollout_044/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_044/agentview/0009.png", "metadata/rollout_044/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_044/agentview/0010.png", "metadata/rollout_044/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_044/agentview/0011.png", "metadata/rollout_044/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_044/agentview/0012.png", "metadata/rollout_044/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_044/agentview/0013.png", "metadata/rollout_044/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_044/agentview/0014.png", "metadata/rollout_044/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_044/agentview/0015.png", "metadata/rollout_044/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_044/agentview/0016.png", "metadata/rollout_044/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_044/agentview/0017.png", "metadata/rollout_044/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_044/agentview/0018.png", "metadata/rollout_044/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_044/agentview/0019.png", "metadata/rollout_044/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_044/agentview/0020.png", "metadata/rollout_044/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0021.png", "metadata/rollout_044/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0022.png", "metadata/rollout_044/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0023.png", "metadata/rollout_044/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0024.png", "metadata/rollout_044/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0025.png", "metadata/rollout_044/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0026.png", "metadata/rollout_044/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_044/agentview/0027.png", "metadata/rollout_044/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0028.png", "metadata/rollout_044/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0029.png", "metadata/rollout_044/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0030.png", "metadata/rollout_044/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0031.png", "metadata/rollout_044/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0032.png", "metadata/rollout_044/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0033.png", "metadata/rollout_044/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0034.png", "metadata/rollout_044/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0035.png", "metadata/rollout_044/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_044/agentview/0036.png", "metadata/rollout_044/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_044/agentview/0037.png", "metadata/rollout_044/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_044/agentview/0038.png", "metadata/rollout_044/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_044/agentview/0039.png", "metadata/rollout_044/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_044/agentview/0040.png", "metadata/rollout_044/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_044/agentview/0041.png", "metadata/rollout_044/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_044/agentview/0042.png", "metadata/rollout_044/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_044/agentview/0043.png", "metadata/rollout_044/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_044/agentview/0044.png", "metadata/rollout_044/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_044/agentview/0045.png", "metadata/rollout_044/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0046.png", "metadata/rollout_044/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0047.png", "metadata/rollout_044/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0048.png", "metadata/rollout_044/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_044/agentview/0049.png", "metadata/rollout_044/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_044/agentview/0050.png", "metadata/rollout_044/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0051.png", "metadata/rollout_044/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_044/agentview/0052.png", "metadata/rollout_044/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_044/agentview/0052.png", "metadata/rollout_044/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0000.png", "metadata/rollout_045/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0001.png", "metadata/rollout_045/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0002.png", "metadata/rollout_045/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0003.png", "metadata/rollout_045/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0004.png", "metadata/rollout_045/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0005.png", "metadata/rollout_045/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0006.png", "metadata/rollout_045/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0007.png", "metadata/rollout_045/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0008.png", "metadata/rollout_045/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_045/agentview/0009.png", "metadata/rollout_045/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_045/agentview/0010.png", "metadata/rollout_045/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_045/agentview/0011.png", "metadata/rollout_045/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_045/agentview/0012.png", "metadata/rollout_045/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_045/agentview/0013.png", "metadata/rollout_045/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0014.png", "metadata/rollout_045/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0015.png", "metadata/rollout_045/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0016.png", "metadata/rollout_045/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0017.png", "metadata/rollout_045/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0018.png", "metadata/rollout_045/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0019.png", "metadata/rollout_045/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_045/agentview/0020.png", "metadata/rollout_045/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0021.png", "metadata/rollout_045/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0022.png", "metadata/rollout_045/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0023.png", "metadata/rollout_045/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0024.png", "metadata/rollout_045/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0025.png", "metadata/rollout_045/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0026.png", "metadata/rollout_045/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0027.png", "metadata/rollout_045/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0028.png", "metadata/rollout_045/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0029.png", "metadata/rollout_045/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0030.png", "metadata/rollout_045/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0031.png", "metadata/rollout_045/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0032.png", "metadata/rollout_045/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0033.png", "metadata/rollout_045/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0034.png", "metadata/rollout_045/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0035.png", "metadata/rollout_045/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0036.png", "metadata/rollout_045/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0037.png", "metadata/rollout_045/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0038.png", "metadata/rollout_045/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_045/agentview/0039.png", "metadata/rollout_045/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0040.png", "metadata/rollout_045/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0041.png", "metadata/rollout_045/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0042.png", "metadata/rollout_045/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_045/agentview/0043.png", "metadata/rollout_045/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_045/agentview/0044.png", "metadata/rollout_045/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0045.png", "metadata/rollout_045/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_045/agentview/0046.png", "metadata/rollout_045/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_045/agentview/0046.png", "metadata/rollout_045/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0000.png", "metadata/rollout_046/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0001.png", "metadata/rollout_046/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0002.png", "metadata/rollout_046/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0003.png", "metadata/rollout_046/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0004.png", "metadata/rollout_046/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0005.png", "metadata/rollout_046/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0006.png", "metadata/rollout_046/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0007.png", "metadata/rollout_046/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0008.png", "metadata/rollout_046/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0009.png", "metadata/rollout_046/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_046/agentview/0010.png", "metadata/rollout_046/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0011.png", "metadata/rollout_046/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0012.png", "metadata/rollout_046/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0013.png", "metadata/rollout_046/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0014.png", "metadata/rollout_046/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0015.png", "metadata/rollout_046/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0016.png", "metadata/rollout_046/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0017.png", "metadata/rollout_046/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0018.png", "metadata/rollout_046/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0019.png", "metadata/rollout_046/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_046/agentview/0020.png", "metadata/rollout_046/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_046/agentview/0021.png", "metadata/rollout_046/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_046/agentview/0022.png", "metadata/rollout_046/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_046/agentview/0023.png", "metadata/rollout_046/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_046/agentview/0024.png", "metadata/rollout_046/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_046/agentview/0025.png", "metadata/rollout_046/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0026.png", "metadata/rollout_046/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0027.png", "metadata/rollout_046/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0028.png", "metadata/rollout_046/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0029.png", "metadata/rollout_046/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0030.png", "metadata/rollout_046/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_046/agentview/0031.png", "metadata/rollout_046/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_046/agentview/0032.png", "metadata/rollout_046/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_046/agentview/0033.png", "metadata/rollout_046/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_046/agentview/0034.png", "metadata/rollout_046/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_046/agentview/0035.png", "metadata/rollout_046/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_046/agentview/0036.png", "metadata/rollout_046/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_046/agentview/0037.png", "metadata/rollout_046/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_046/agentview/0038.png", "metadata/rollout_046/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0039.png", "metadata/rollout_046/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0040.png", "metadata/rollout_046/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0041.png", "metadata/rollout_046/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0042.png", "metadata/rollout_046/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0043.png", "metadata/rollout_046/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0044.png", "metadata/rollout_046/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0045.png", "metadata/rollout_046/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0046.png", "metadata/rollout_046/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0047.png", "metadata/rollout_046/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0048.png", "metadata/rollout_046/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0049.png", "metadata/rollout_046/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0050.png", "metadata/rollout_046/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0051.png", "metadata/rollout_046/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0052.png", "metadata/rollout_046/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_046/agentview/0053.png", "metadata/rollout_046/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_046/agentview/0054.png", "metadata/rollout_046/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_046/agentview/0055.png", "metadata/rollout_046/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_046/agentview/0056.png", "metadata/rollout_046/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_046/agentview/0057.png", "metadata/rollout_046/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_046/agentview/0058.png", "metadata/rollout_046/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0059.png", "metadata/rollout_046/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0060.png", "metadata/rollout_046/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_046/agentview/0061.png", "metadata/rollout_046/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_046/agentview/0062.png", "metadata/rollout_046/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_046/agentview/0063.png", "metadata/rollout_046/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_046/agentview/0064.png", "metadata/rollout_046/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_046/agentview/0064.png", "metadata/rollout_046/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0000.png", "metadata/rollout_047/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0001.png", "metadata/rollout_047/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0002.png", "metadata/rollout_047/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0003.png", "metadata/rollout_047/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0004.png", "metadata/rollout_047/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0005.png", "metadata/rollout_047/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0006.png", "metadata/rollout_047/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0007.png", "metadata/rollout_047/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0008.png", "metadata/rollout_047/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_047/agentview/0009.png", "metadata/rollout_047/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0010.png", "metadata/rollout_047/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0011.png", "metadata/rollout_047/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0012.png", "metadata/rollout_047/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0013.png", "metadata/rollout_047/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0014.png", "metadata/rollout_047/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0015.png", "metadata/rollout_047/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0016.png", "metadata/rollout_047/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0017.png", "metadata/rollout_047/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0018.png", "metadata/rollout_047/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_047/agentview/0019.png", "metadata/rollout_047/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_047/agentview/0020.png", "metadata/rollout_047/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_047/agentview/0021.png", "metadata/rollout_047/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_047/agentview/0022.png", "metadata/rollout_047/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_047/agentview/0023.png", "metadata/rollout_047/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_047/agentview/0024.png", "metadata/rollout_047/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_047/agentview/0025.png", "metadata/rollout_047/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_047/agentview/0026.png", "metadata/rollout_047/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_047/agentview/0027.png", "metadata/rollout_047/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0028.png", "metadata/rollout_047/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0029.png", "metadata/rollout_047/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0030.png", "metadata/rollout_047/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0031.png", "metadata/rollout_047/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0032.png", "metadata/rollout_047/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0033.png", "metadata/rollout_047/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_047/agentview/0034.png", "metadata/rollout_047/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0035.png", "metadata/rollout_047/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0036.png", "metadata/rollout_047/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0037.png", "metadata/rollout_047/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0038.png", "metadata/rollout_047/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0039.png", "metadata/rollout_047/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0040.png", "metadata/rollout_047/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0041.png", "metadata/rollout_047/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0042.png", "metadata/rollout_047/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0043.png", "metadata/rollout_047/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0044.png", "metadata/rollout_047/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0045.png", "metadata/rollout_047/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0046.png", "metadata/rollout_047/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0047.png", "metadata/rollout_047/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0048.png", "metadata/rollout_047/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0049.png", "metadata/rollout_047/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0050.png", "metadata/rollout_047/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0051.png", "metadata/rollout_047/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0052.png", "metadata/rollout_047/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0053.png", "metadata/rollout_047/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0054.png", "metadata/rollout_047/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0055.png", "metadata/rollout_047/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0056.png", "metadata/rollout_047/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0057.png", "metadata/rollout_047/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_047/agentview/0058.png", "metadata/rollout_047/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0059.png", "metadata/rollout_047/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0060.png", "metadata/rollout_047/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0061.png", "metadata/rollout_047/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_047/agentview/0062.png", "metadata/rollout_047/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_047/agentview/0063.png", "metadata/rollout_047/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0064.png", "metadata/rollout_047/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_047/agentview/0065.png", "metadata/rollout_047/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_047/agentview/0065.png", "metadata/rollout_047/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0000.png", "metadata/rollout_048/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0001.png", "metadata/rollout_048/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0002.png", "metadata/rollout_048/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0003.png", "metadata/rollout_048/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0004.png", "metadata/rollout_048/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0005.png", "metadata/rollout_048/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0006.png", "metadata/rollout_048/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0007.png", "metadata/rollout_048/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0008.png", "metadata/rollout_048/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_048/agentview/0009.png", "metadata/rollout_048/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0010.png", "metadata/rollout_048/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0011.png", "metadata/rollout_048/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0012.png", "metadata/rollout_048/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0013.png", "metadata/rollout_048/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0014.png", "metadata/rollout_048/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0015.png", "metadata/rollout_048/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0016.png", "metadata/rollout_048/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0017.png", "metadata/rollout_048/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0018.png", "metadata/rollout_048/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_048/agentview/0019.png", "metadata/rollout_048/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_048/agentview/0020.png", "metadata/rollout_048/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_048/agentview/0021.png", "metadata/rollout_048/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0022.png", "metadata/rollout_048/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0023.png", "metadata/rollout_048/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0024.png", "metadata/rollout_048/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0025.png", "metadata/rollout_048/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0026.png", "metadata/rollout_048/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0027.png", "metadata/rollout_048/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_048/agentview/0028.png", "metadata/rollout_048/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0029.png", "metadata/rollout_048/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0030.png", "metadata/rollout_048/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0031.png", "metadata/rollout_048/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0032.png", "metadata/rollout_048/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0033.png", "metadata/rollout_048/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0034.png", "metadata/rollout_048/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0035.png", "metadata/rollout_048/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0036.png", "metadata/rollout_048/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0037.png", "metadata/rollout_048/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0038.png", "metadata/rollout_048/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0039.png", "metadata/rollout_048/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0040.png", "metadata/rollout_048/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0041.png", "metadata/rollout_048/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0042.png", "metadata/rollout_048/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0043.png", "metadata/rollout_048/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0044.png", "metadata/rollout_048/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0045.png", "metadata/rollout_048/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_048/agentview/0046.png", "metadata/rollout_048/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0047.png", "metadata/rollout_048/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0048.png", "metadata/rollout_048/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0049.png", "metadata/rollout_048/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_048/agentview/0050.png", "metadata/rollout_048/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_048/agentview/0051.png", "metadata/rollout_048/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0052.png", "metadata/rollout_048/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_048/agentview/0053.png", "metadata/rollout_048/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_048/agentview/0053.png", "metadata/rollout_048/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0000.png", "metadata/rollout_049/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0001.png", "metadata/rollout_049/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0002.png", "metadata/rollout_049/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0003.png", "metadata/rollout_049/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0004.png", "metadata/rollout_049/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0005.png", "metadata/rollout_049/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0006.png", "metadata/rollout_049/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0007.png", "metadata/rollout_049/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0008.png", "metadata/rollout_049/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_049/agentview/0009.png", "metadata/rollout_049/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0010.png", "metadata/rollout_049/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0011.png", "metadata/rollout_049/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0012.png", "metadata/rollout_049/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0013.png", "metadata/rollout_049/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0014.png", "metadata/rollout_049/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0015.png", "metadata/rollout_049/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0016.png", "metadata/rollout_049/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0017.png", "metadata/rollout_049/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0018.png", "metadata/rollout_049/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_049/agentview/0019.png", "metadata/rollout_049/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_049/agentview/0020.png", "metadata/rollout_049/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_049/agentview/0021.png", "metadata/rollout_049/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_049/agentview/0022.png", "metadata/rollout_049/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_049/agentview/0023.png", "metadata/rollout_049/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_049/agentview/0024.png", "metadata/rollout_049/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0025.png", "metadata/rollout_049/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0026.png", "metadata/rollout_049/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0027.png", "metadata/rollout_049/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0028.png", "metadata/rollout_049/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0029.png", "metadata/rollout_049/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0030.png", "metadata/rollout_049/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_049/agentview/0031.png", "metadata/rollout_049/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0032.png", "metadata/rollout_049/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0033.png", "metadata/rollout_049/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0034.png", "metadata/rollout_049/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0035.png", "metadata/rollout_049/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0036.png", "metadata/rollout_049/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0037.png", "metadata/rollout_049/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0038.png", "metadata/rollout_049/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0039.png", "metadata/rollout_049/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0040.png", "metadata/rollout_049/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0041.png", "metadata/rollout_049/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0042.png", "metadata/rollout_049/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0043.png", "metadata/rollout_049/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0044.png", "metadata/rollout_049/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0045.png", "metadata/rollout_049/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0046.png", "metadata/rollout_049/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0047.png", "metadata/rollout_049/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0048.png", "metadata/rollout_049/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0049.png", "metadata/rollout_049/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_049/agentview/0050.png", "metadata/rollout_049/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_049/agentview/0051.png", "metadata/rollout_049/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0052.png", "metadata/rollout_049/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0053.png", "metadata/rollout_049/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0054.png", "metadata/rollout_049/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_049/agentview/0055.png", "metadata/rollout_049/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_049/agentview/0056.png", "metadata/rollout_049/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0057.png", "metadata/rollout_049/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_049/agentview/0058.png", "metadata/rollout_049/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the green block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_049/agentview/0058.png", "metadata/rollout_049/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0000.png", "metadata/rollout_050/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0001.png", "metadata/rollout_050/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0002.png", "metadata/rollout_050/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0003.png", "metadata/rollout_050/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0004.png", "metadata/rollout_050/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0005.png", "metadata/rollout_050/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0006.png", "metadata/rollout_050/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0007.png", "metadata/rollout_050/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0008.png", "metadata/rollout_050/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_050/agentview/0009.png", "metadata/rollout_050/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_050/agentview/0010.png", "metadata/rollout_050/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_050/agentview/0011.png", "metadata/rollout_050/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_050/agentview/0012.png", "metadata/rollout_050/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_050/agentview/0013.png", "metadata/rollout_050/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_050/agentview/0014.png", "metadata/rollout_050/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_050/agentview/0015.png", "metadata/rollout_050/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0016.png", "metadata/rollout_050/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0017.png", "metadata/rollout_050/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0018.png", "metadata/rollout_050/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0019.png", "metadata/rollout_050/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0020.png", "metadata/rollout_050/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0021.png", "metadata/rollout_050/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_050/agentview/0022.png", "metadata/rollout_050/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0023.png", "metadata/rollout_050/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0024.png", "metadata/rollout_050/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0025.png", "metadata/rollout_050/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0026.png", "metadata/rollout_050/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0027.png", "metadata/rollout_050/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0028.png", "metadata/rollout_050/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0029.png", "metadata/rollout_050/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0030.png", "metadata/rollout_050/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_050/agentview/0031.png", "metadata/rollout_050/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_050/agentview/0032.png", "metadata/rollout_050/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_050/agentview/0033.png", "metadata/rollout_050/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_050/agentview/0034.png", "metadata/rollout_050/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_050/agentview/0035.png", "metadata/rollout_050/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_050/agentview/0036.png", "metadata/rollout_050/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_050/agentview/0037.png", "metadata/rollout_050/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_050/agentview/0038.png", "metadata/rollout_050/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_050/agentview/0039.png", "metadata/rollout_050/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0040.png", "metadata/rollout_050/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0041.png", "metadata/rollout_050/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0042.png", "metadata/rollout_050/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_050/agentview/0043.png", "metadata/rollout_050/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_050/agentview/0044.png", "metadata/rollout_050/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0045.png", "metadata/rollout_050/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_050/agentview/0046.png", "metadata/rollout_050/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_050/agentview/0046.png", "metadata/rollout_050/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0000.png", "metadata/rollout_051/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0001.png", "metadata/rollout_051/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0002.png", "metadata/rollout_051/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0003.png", "metadata/rollout_051/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0004.png", "metadata/rollout_051/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0005.png", "metadata/rollout_051/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0006.png", "metadata/rollout_051/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0007.png", "metadata/rollout_051/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0008.png", "metadata/rollout_051/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_051/agentview/0009.png", "metadata/rollout_051/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_051/agentview/0010.png", "metadata/rollout_051/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_051/agentview/0011.png", "metadata/rollout_051/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_051/agentview/0012.png", "metadata/rollout_051/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_051/agentview/0013.png", "metadata/rollout_051/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_051/agentview/0014.png", "metadata/rollout_051/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_051/agentview/0015.png", "metadata/rollout_051/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_051/agentview/0016.png", "metadata/rollout_051/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_051/agentview/0017.png", "metadata/rollout_051/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0018.png", "metadata/rollout_051/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0019.png", "metadata/rollout_051/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0020.png", "metadata/rollout_051/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0021.png", "metadata/rollout_051/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0022.png", "metadata/rollout_051/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0023.png", "metadata/rollout_051/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_051/agentview/0024.png", "metadata/rollout_051/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0025.png", "metadata/rollout_051/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0026.png", "metadata/rollout_051/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0027.png", "metadata/rollout_051/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0028.png", "metadata/rollout_051/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0029.png", "metadata/rollout_051/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0030.png", "metadata/rollout_051/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0031.png", "metadata/rollout_051/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0032.png", "metadata/rollout_051/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0033.png", "metadata/rollout_051/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0034.png", "metadata/rollout_051/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0035.png", "metadata/rollout_051/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0036.png", "metadata/rollout_051/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0037.png", "metadata/rollout_051/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0038.png", "metadata/rollout_051/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0039.png", "metadata/rollout_051/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0040.png", "metadata/rollout_051/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0041.png", "metadata/rollout_051/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0042.png", "metadata/rollout_051/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_051/agentview/0043.png", "metadata/rollout_051/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_051/agentview/0044.png", "metadata/rollout_051/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_051/agentview/0045.png", "metadata/rollout_051/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_051/agentview/0046.png", "metadata/rollout_051/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0047.png", "metadata/rollout_051/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0048.png", "metadata/rollout_051/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0049.png", "metadata/rollout_051/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_051/agentview/0050.png", "metadata/rollout_051/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_051/agentview/0051.png", "metadata/rollout_051/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0052.png", "metadata/rollout_051/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_051/agentview/0053.png", "metadata/rollout_051/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_051/agentview/0053.png", "metadata/rollout_051/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0000.png", "metadata/rollout_052/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0001.png", "metadata/rollout_052/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0002.png", "metadata/rollout_052/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0003.png", "metadata/rollout_052/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0004.png", "metadata/rollout_052/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0005.png", "metadata/rollout_052/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0006.png", "metadata/rollout_052/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0007.png", "metadata/rollout_052/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0008.png", "metadata/rollout_052/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_052/agentview/0009.png", "metadata/rollout_052/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_052/agentview/0010.png", "metadata/rollout_052/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_052/agentview/0011.png", "metadata/rollout_052/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_052/agentview/0012.png", "metadata/rollout_052/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_052/agentview/0013.png", "metadata/rollout_052/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_052/agentview/0014.png", "metadata/rollout_052/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_052/agentview/0015.png", "metadata/rollout_052/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_052/agentview/0016.png", "metadata/rollout_052/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_052/agentview/0017.png", "metadata/rollout_052/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_052/agentview/0018.png", "metadata/rollout_052/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_052/agentview/0019.png", "metadata/rollout_052/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_052/agentview/0020.png", "metadata/rollout_052/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_052/agentview/0021.png", "metadata/rollout_052/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_052/agentview/0022.png", "metadata/rollout_052/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_052/agentview/0023.png", "metadata/rollout_052/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_052/agentview/0024.png", "metadata/rollout_052/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_052/agentview/0025.png", "metadata/rollout_052/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0026.png", "metadata/rollout_052/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0027.png", "metadata/rollout_052/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0028.png", "metadata/rollout_052/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0029.png", "metadata/rollout_052/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0030.png", "metadata/rollout_052/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0031.png", "metadata/rollout_052/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_052/agentview/0032.png", "metadata/rollout_052/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0033.png", "metadata/rollout_052/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0034.png", "metadata/rollout_052/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0035.png", "metadata/rollout_052/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0036.png", "metadata/rollout_052/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0037.png", "metadata/rollout_052/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0038.png", "metadata/rollout_052/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0039.png", "metadata/rollout_052/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0040.png", "metadata/rollout_052/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_052/agentview/0041.png", "metadata/rollout_052/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_052/agentview/0042.png", "metadata/rollout_052/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_052/agentview/0043.png", "metadata/rollout_052/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0044.png", "metadata/rollout_052/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0045.png", "metadata/rollout_052/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0046.png", "metadata/rollout_052/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0047.png", "metadata/rollout_052/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0048.png", "metadata/rollout_052/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0049.png", "metadata/rollout_052/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0050.png", "metadata/rollout_052/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0051.png", "metadata/rollout_052/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0052.png", "metadata/rollout_052/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0053.png", "metadata/rollout_052/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0054.png", "metadata/rollout_052/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0055.png", "metadata/rollout_052/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_052/agentview/0056.png", "metadata/rollout_052/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0057.png", "metadata/rollout_052/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0058.png", "metadata/rollout_052/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0059.png", "metadata/rollout_052/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_052/agentview/0060.png", "metadata/rollout_052/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_052/agentview/0061.png", "metadata/rollout_052/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0062.png", "metadata/rollout_052/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_052/agentview/0063.png", "metadata/rollout_052/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_052/agentview/0063.png", "metadata/rollout_052/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0000.png", "metadata/rollout_053/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0001.png", "metadata/rollout_053/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0002.png", "metadata/rollout_053/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0003.png", "metadata/rollout_053/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0004.png", "metadata/rollout_053/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0005.png", "metadata/rollout_053/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0006.png", "metadata/rollout_053/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0007.png", "metadata/rollout_053/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0008.png", "metadata/rollout_053/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0009.png", "metadata/rollout_053/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0010.png", "metadata/rollout_053/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0011.png", "metadata/rollout_053/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0012.png", "metadata/rollout_053/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0013.png", "metadata/rollout_053/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0014.png", "metadata/rollout_053/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0015.png", "metadata/rollout_053/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0016.png", "metadata/rollout_053/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0017.png", "metadata/rollout_053/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_053/agentview/0018.png", "metadata/rollout_053/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_053/agentview/0019.png", "metadata/rollout_053/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_053/agentview/0020.png", "metadata/rollout_053/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_053/agentview/0021.png", "metadata/rollout_053/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_053/agentview/0022.png", "metadata/rollout_053/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_053/agentview/0023.png", "metadata/rollout_053/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_053/agentview/0024.png", "metadata/rollout_053/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0025.png", "metadata/rollout_053/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0026.png", "metadata/rollout_053/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0027.png", "metadata/rollout_053/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0028.png", "metadata/rollout_053/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0029.png", "metadata/rollout_053/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0030.png", "metadata/rollout_053/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_053/agentview/0031.png", "metadata/rollout_053/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0032.png", "metadata/rollout_053/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0033.png", "metadata/rollout_053/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0034.png", "metadata/rollout_053/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0035.png", "metadata/rollout_053/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0036.png", "metadata/rollout_053/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0037.png", "metadata/rollout_053/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0038.png", "metadata/rollout_053/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0039.png", "metadata/rollout_053/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0040.png", "metadata/rollout_053/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0041.png", "metadata/rollout_053/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0042.png", "metadata/rollout_053/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0043.png", "metadata/rollout_053/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0044.png", "metadata/rollout_053/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0045.png", "metadata/rollout_053/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0046.png", "metadata/rollout_053/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0047.png", "metadata/rollout_053/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0048.png", "metadata/rollout_053/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0049.png", "metadata/rollout_053/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0050.png", "metadata/rollout_053/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0051.png", "metadata/rollout_053/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0052.png", "metadata/rollout_053/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_053/agentview/0053.png", "metadata/rollout_053/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_053/agentview/0054.png", "metadata/rollout_053/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0055.png", "metadata/rollout_053/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0056.png", "metadata/rollout_053/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0057.png", "metadata/rollout_053/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_053/agentview/0058.png", "metadata/rollout_053/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_053/agentview/0059.png", "metadata/rollout_053/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0060.png", "metadata/rollout_053/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_053/agentview/0061.png", "metadata/rollout_053/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_053/agentview/0061.png", "metadata/rollout_053/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0000.png", "metadata/rollout_054/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0001.png", "metadata/rollout_054/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0002.png", "metadata/rollout_054/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0003.png", "metadata/rollout_054/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0004.png", "metadata/rollout_054/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0005.png", "metadata/rollout_054/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0006.png", "metadata/rollout_054/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0007.png", "metadata/rollout_054/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0008.png", "metadata/rollout_054/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_054/agentview/0009.png", "metadata/rollout_054/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_054/agentview/0010.png", "metadata/rollout_054/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_054/agentview/0011.png", "metadata/rollout_054/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_054/agentview/0012.png", "metadata/rollout_054/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_054/agentview/0013.png", "metadata/rollout_054/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_054/agentview/0014.png", "metadata/rollout_054/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_054/agentview/0015.png", "metadata/rollout_054/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_054/agentview/0016.png", "metadata/rollout_054/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_054/agentview/0017.png", "metadata/rollout_054/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_054/agentview/0018.png", "metadata/rollout_054/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_054/agentview/0019.png", "metadata/rollout_054/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_054/agentview/0020.png", "metadata/rollout_054/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0021.png", "metadata/rollout_054/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0022.png", "metadata/rollout_054/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0023.png", "metadata/rollout_054/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0024.png", "metadata/rollout_054/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0025.png", "metadata/rollout_054/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0026.png", "metadata/rollout_054/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_054/agentview/0027.png", "metadata/rollout_054/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0028.png", "metadata/rollout_054/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0029.png", "metadata/rollout_054/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0030.png", "metadata/rollout_054/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0031.png", "metadata/rollout_054/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0032.png", "metadata/rollout_054/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0033.png", "metadata/rollout_054/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0034.png", "metadata/rollout_054/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0035.png", "metadata/rollout_054/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_054/agentview/0036.png", "metadata/rollout_054/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_054/agentview/0037.png", "metadata/rollout_054/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_054/agentview/0038.png", "metadata/rollout_054/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_054/agentview/0039.png", "metadata/rollout_054/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_054/agentview/0040.png", "metadata/rollout_054/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_054/agentview/0041.png", "metadata/rollout_054/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_054/agentview/0042.png", "metadata/rollout_054/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_054/agentview/0043.png", "metadata/rollout_054/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_054/agentview/0044.png", "metadata/rollout_054/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_054/agentview/0045.png", "metadata/rollout_054/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0046.png", "metadata/rollout_054/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0047.png", "metadata/rollout_054/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0048.png", "metadata/rollout_054/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_054/agentview/0049.png", "metadata/rollout_054/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_054/agentview/0050.png", "metadata/rollout_054/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0051.png", "metadata/rollout_054/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_054/agentview/0052.png", "metadata/rollout_054/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_054/agentview/0052.png", "metadata/rollout_054/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0000.png", "metadata/rollout_055/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0001.png", "metadata/rollout_055/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0002.png", "metadata/rollout_055/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0003.png", "metadata/rollout_055/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0004.png", "metadata/rollout_055/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0005.png", "metadata/rollout_055/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0006.png", "metadata/rollout_055/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0007.png", "metadata/rollout_055/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0008.png", "metadata/rollout_055/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_055/agentview/0009.png", "metadata/rollout_055/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_055/agentview/0010.png", "metadata/rollout_055/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_055/agentview/0011.png", "metadata/rollout_055/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_055/agentview/0012.png", "metadata/rollout_055/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_055/agentview/0013.png", "metadata/rollout_055/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0014.png", "metadata/rollout_055/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0015.png", "metadata/rollout_055/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0016.png", "metadata/rollout_055/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0017.png", "metadata/rollout_055/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0018.png", "metadata/rollout_055/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0019.png", "metadata/rollout_055/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_055/agentview/0020.png", "metadata/rollout_055/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0021.png", "metadata/rollout_055/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0022.png", "metadata/rollout_055/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0023.png", "metadata/rollout_055/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0024.png", "metadata/rollout_055/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0025.png", "metadata/rollout_055/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0026.png", "metadata/rollout_055/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0027.png", "metadata/rollout_055/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0028.png", "metadata/rollout_055/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0029.png", "metadata/rollout_055/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0030.png", "metadata/rollout_055/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0031.png", "metadata/rollout_055/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0032.png", "metadata/rollout_055/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0033.png", "metadata/rollout_055/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0034.png", "metadata/rollout_055/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0035.png", "metadata/rollout_055/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0036.png", "metadata/rollout_055/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0037.png", "metadata/rollout_055/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0038.png", "metadata/rollout_055/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_055/agentview/0039.png", "metadata/rollout_055/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0040.png", "metadata/rollout_055/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0041.png", "metadata/rollout_055/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0042.png", "metadata/rollout_055/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_055/agentview/0043.png", "metadata/rollout_055/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_055/agentview/0044.png", "metadata/rollout_055/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0045.png", "metadata/rollout_055/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_055/agentview/0046.png", "metadata/rollout_055/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_055/agentview/0046.png", "metadata/rollout_055/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0000.png", "metadata/rollout_056/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0001.png", "metadata/rollout_056/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0002.png", "metadata/rollout_056/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0003.png", "metadata/rollout_056/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0004.png", "metadata/rollout_056/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0005.png", "metadata/rollout_056/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0006.png", "metadata/rollout_056/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0007.png", "metadata/rollout_056/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0008.png", "metadata/rollout_056/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0009.png", "metadata/rollout_056/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_056/agentview/0010.png", "metadata/rollout_056/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0011.png", "metadata/rollout_056/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0012.png", "metadata/rollout_056/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0013.png", "metadata/rollout_056/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0014.png", "metadata/rollout_056/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0015.png", "metadata/rollout_056/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0016.png", "metadata/rollout_056/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0017.png", "metadata/rollout_056/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0018.png", "metadata/rollout_056/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0019.png", "metadata/rollout_056/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_056/agentview/0020.png", "metadata/rollout_056/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_056/agentview/0021.png", "metadata/rollout_056/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_056/agentview/0022.png", "metadata/rollout_056/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_056/agentview/0023.png", "metadata/rollout_056/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_056/agentview/0024.png", "metadata/rollout_056/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_056/agentview/0025.png", "metadata/rollout_056/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0026.png", "metadata/rollout_056/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0027.png", "metadata/rollout_056/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0028.png", "metadata/rollout_056/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0029.png", "metadata/rollout_056/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0030.png", "metadata/rollout_056/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_056/agentview/0031.png", "metadata/rollout_056/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_056/agentview/0032.png", "metadata/rollout_056/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_056/agentview/0033.png", "metadata/rollout_056/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_056/agentview/0034.png", "metadata/rollout_056/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_056/agentview/0035.png", "metadata/rollout_056/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_056/agentview/0036.png", "metadata/rollout_056/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_056/agentview/0037.png", "metadata/rollout_056/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_056/agentview/0038.png", "metadata/rollout_056/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0039.png", "metadata/rollout_056/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0040.png", "metadata/rollout_056/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0041.png", "metadata/rollout_056/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0042.png", "metadata/rollout_056/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0043.png", "metadata/rollout_056/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0044.png", "metadata/rollout_056/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0045.png", "metadata/rollout_056/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0046.png", "metadata/rollout_056/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0047.png", "metadata/rollout_056/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0048.png", "metadata/rollout_056/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0049.png", "metadata/rollout_056/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0050.png", "metadata/rollout_056/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0051.png", "metadata/rollout_056/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0052.png", "metadata/rollout_056/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_056/agentview/0053.png", "metadata/rollout_056/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_056/agentview/0054.png", "metadata/rollout_056/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_056/agentview/0055.png", "metadata/rollout_056/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_056/agentview/0056.png", "metadata/rollout_056/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_056/agentview/0057.png", "metadata/rollout_056/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_056/agentview/0058.png", "metadata/rollout_056/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0059.png", "metadata/rollout_056/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0060.png", "metadata/rollout_056/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_056/agentview/0061.png", "metadata/rollout_056/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_056/agentview/0062.png", "metadata/rollout_056/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_056/agentview/0063.png", "metadata/rollout_056/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_056/agentview/0064.png", "metadata/rollout_056/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_056/agentview/0064.png", "metadata/rollout_056/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0000.png", "metadata/rollout_057/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0001.png", "metadata/rollout_057/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0002.png", "metadata/rollout_057/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0003.png", "metadata/rollout_057/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0004.png", "metadata/rollout_057/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0005.png", "metadata/rollout_057/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0006.png", "metadata/rollout_057/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0007.png", "metadata/rollout_057/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0008.png", "metadata/rollout_057/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_057/agentview/0009.png", "metadata/rollout_057/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0010.png", "metadata/rollout_057/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0011.png", "metadata/rollout_057/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0012.png", "metadata/rollout_057/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0013.png", "metadata/rollout_057/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0014.png", "metadata/rollout_057/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0015.png", "metadata/rollout_057/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0016.png", "metadata/rollout_057/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0017.png", "metadata/rollout_057/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0018.png", "metadata/rollout_057/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_057/agentview/0019.png", "metadata/rollout_057/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_057/agentview/0020.png", "metadata/rollout_057/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_057/agentview/0021.png", "metadata/rollout_057/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_057/agentview/0022.png", "metadata/rollout_057/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_057/agentview/0023.png", "metadata/rollout_057/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_057/agentview/0024.png", "metadata/rollout_057/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_057/agentview/0025.png", "metadata/rollout_057/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_057/agentview/0026.png", "metadata/rollout_057/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_057/agentview/0027.png", "metadata/rollout_057/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0028.png", "metadata/rollout_057/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0029.png", "metadata/rollout_057/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0030.png", "metadata/rollout_057/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0031.png", "metadata/rollout_057/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0032.png", "metadata/rollout_057/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0033.png", "metadata/rollout_057/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_057/agentview/0034.png", "metadata/rollout_057/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0035.png", "metadata/rollout_057/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0036.png", "metadata/rollout_057/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0037.png", "metadata/rollout_057/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0038.png", "metadata/rollout_057/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0039.png", "metadata/rollout_057/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0040.png", "metadata/rollout_057/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0041.png", "metadata/rollout_057/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0042.png", "metadata/rollout_057/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0043.png", "metadata/rollout_057/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0044.png", "metadata/rollout_057/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0045.png", "metadata/rollout_057/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0046.png", "metadata/rollout_057/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0047.png", "metadata/rollout_057/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0048.png", "metadata/rollout_057/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0049.png", "metadata/rollout_057/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0050.png", "metadata/rollout_057/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0051.png", "metadata/rollout_057/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0052.png", "metadata/rollout_057/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0053.png", "metadata/rollout_057/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0054.png", "metadata/rollout_057/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0055.png", "metadata/rollout_057/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0056.png", "metadata/rollout_057/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0057.png", "metadata/rollout_057/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_057/agentview/0058.png", "metadata/rollout_057/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0059.png", "metadata/rollout_057/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0060.png", "metadata/rollout_057/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0061.png", "metadata/rollout_057/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_057/agentview/0062.png", "metadata/rollout_057/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_057/agentview/0063.png", "metadata/rollout_057/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0064.png", "metadata/rollout_057/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_057/agentview/0065.png", "metadata/rollout_057/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_057/agentview/0065.png", "metadata/rollout_057/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0000.png", "metadata/rollout_058/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0001.png", "metadata/rollout_058/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0002.png", "metadata/rollout_058/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0003.png", "metadata/rollout_058/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0004.png", "metadata/rollout_058/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0005.png", "metadata/rollout_058/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0006.png", "metadata/rollout_058/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0007.png", "metadata/rollout_058/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0008.png", "metadata/rollout_058/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_058/agentview/0009.png", "metadata/rollout_058/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0010.png", "metadata/rollout_058/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0011.png", "metadata/rollout_058/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0012.png", "metadata/rollout_058/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0013.png", "metadata/rollout_058/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0014.png", "metadata/rollout_058/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0015.png", "metadata/rollout_058/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0016.png", "metadata/rollout_058/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0017.png", "metadata/rollout_058/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0018.png", "metadata/rollout_058/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_058/agentview/0019.png", "metadata/rollout_058/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_058/agentview/0020.png", "metadata/rollout_058/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_058/agentview/0021.png", "metadata/rollout_058/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0022.png", "metadata/rollout_058/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0023.png", "metadata/rollout_058/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0024.png", "metadata/rollout_058/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0025.png", "metadata/rollout_058/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0026.png", "metadata/rollout_058/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0027.png", "metadata/rollout_058/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_058/agentview/0028.png", "metadata/rollout_058/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0029.png", "metadata/rollout_058/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0030.png", "metadata/rollout_058/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0031.png", "metadata/rollout_058/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0032.png", "metadata/rollout_058/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0033.png", "metadata/rollout_058/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0034.png", "metadata/rollout_058/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0035.png", "metadata/rollout_058/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0036.png", "metadata/rollout_058/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0037.png", "metadata/rollout_058/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0038.png", "metadata/rollout_058/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0039.png", "metadata/rollout_058/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0040.png", "metadata/rollout_058/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0041.png", "metadata/rollout_058/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0042.png", "metadata/rollout_058/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0043.png", "metadata/rollout_058/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0044.png", "metadata/rollout_058/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0045.png", "metadata/rollout_058/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0046.png", "metadata/rollout_058/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_058/agentview/0047.png", "metadata/rollout_058/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0048.png", "metadata/rollout_058/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0049.png", "metadata/rollout_058/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0050.png", "metadata/rollout_058/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_058/agentview/0051.png", "metadata/rollout_058/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_058/agentview/0052.png", "metadata/rollout_058/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0053.png", "metadata/rollout_058/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_058/agentview/0054.png", "metadata/rollout_058/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_058/agentview/0054.png", "metadata/rollout_058/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0000.png", "metadata/rollout_059/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0001.png", "metadata/rollout_059/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0002.png", "metadata/rollout_059/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0003.png", "metadata/rollout_059/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0004.png", "metadata/rollout_059/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0005.png", "metadata/rollout_059/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0006.png", "metadata/rollout_059/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0007.png", "metadata/rollout_059/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0008.png", "metadata/rollout_059/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_059/agentview/0009.png", "metadata/rollout_059/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0010.png", "metadata/rollout_059/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0011.png", "metadata/rollout_059/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0012.png", "metadata/rollout_059/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0013.png", "metadata/rollout_059/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0014.png", "metadata/rollout_059/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0015.png", "metadata/rollout_059/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0016.png", "metadata/rollout_059/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0017.png", "metadata/rollout_059/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0018.png", "metadata/rollout_059/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_059/agentview/0019.png", "metadata/rollout_059/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_059/agentview/0020.png", "metadata/rollout_059/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_059/agentview/0021.png", "metadata/rollout_059/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_059/agentview/0022.png", "metadata/rollout_059/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_059/agentview/0023.png", "metadata/rollout_059/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_059/agentview/0024.png", "metadata/rollout_059/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0025.png", "metadata/rollout_059/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0026.png", "metadata/rollout_059/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0027.png", "metadata/rollout_059/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0028.png", "metadata/rollout_059/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0029.png", "metadata/rollout_059/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0030.png", "metadata/rollout_059/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_059/agentview/0031.png", "metadata/rollout_059/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0032.png", "metadata/rollout_059/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0033.png", "metadata/rollout_059/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0034.png", "metadata/rollout_059/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0035.png", "metadata/rollout_059/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0036.png", "metadata/rollout_059/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0037.png", "metadata/rollout_059/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0038.png", "metadata/rollout_059/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0039.png", "metadata/rollout_059/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0040.png", "metadata/rollout_059/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0041.png", "metadata/rollout_059/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0042.png", "metadata/rollout_059/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0043.png", "metadata/rollout_059/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0044.png", "metadata/rollout_059/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0045.png", "metadata/rollout_059/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0046.png", "metadata/rollout_059/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0047.png", "metadata/rollout_059/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0048.png", "metadata/rollout_059/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0049.png", "metadata/rollout_059/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_059/agentview/0050.png", "metadata/rollout_059/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_059/agentview/0051.png", "metadata/rollout_059/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0052.png", "metadata/rollout_059/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0053.png", "metadata/rollout_059/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0054.png", "metadata/rollout_059/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_059/agentview/0055.png", "metadata/rollout_059/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_059/agentview/0056.png", "metadata/rollout_059/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0057.png", "metadata/rollout_059/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_059/agentview/0058.png", "metadata/rollout_059/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_059/agentview/0058.png", "metadata/rollout_059/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0000.png", "metadata/rollout_060/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0001.png", "metadata/rollout_060/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0002.png", "metadata/rollout_060/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0003.png", "metadata/rollout_060/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0004.png", "metadata/rollout_060/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0005.png", "metadata/rollout_060/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0006.png", "metadata/rollout_060/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0007.png", "metadata/rollout_060/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0008.png", "metadata/rollout_060/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_060/agentview/0009.png", "metadata/rollout_060/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_060/agentview/0010.png", "metadata/rollout_060/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_060/agentview/0011.png", "metadata/rollout_060/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_060/agentview/0012.png", "metadata/rollout_060/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_060/agentview/0013.png", "metadata/rollout_060/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_060/agentview/0014.png", "metadata/rollout_060/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_060/agentview/0015.png", "metadata/rollout_060/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_060/agentview/0016.png", "metadata/rollout_060/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_060/agentview/0017.png", "metadata/rollout_060/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_060/agentview/0018.png", "metadata/rollout_060/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_060/agentview/0019.png", "metadata/rollout_060/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_060/agentview/0020.png", "metadata/rollout_060/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_060/agentview/0021.png", "metadata/rollout_060/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_060/agentview/0022.png", "metadata/rollout_060/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_060/agentview/0023.png", "metadata/rollout_060/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0024.png", "metadata/rollout_060/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0025.png", "metadata/rollout_060/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0026.png", "metadata/rollout_060/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0027.png", "metadata/rollout_060/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0028.png", "metadata/rollout_060/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0029.png", "metadata/rollout_060/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_060/agentview/0030.png", "metadata/rollout_060/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0031.png", "metadata/rollout_060/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0032.png", "metadata/rollout_060/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0033.png", "metadata/rollout_060/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0034.png", "metadata/rollout_060/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0035.png", "metadata/rollout_060/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0036.png", "metadata/rollout_060/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0037.png", "metadata/rollout_060/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0038.png", "metadata/rollout_060/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0039.png", "metadata/rollout_060/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0040.png", "metadata/rollout_060/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0041.png", "metadata/rollout_060/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0042.png", "metadata/rollout_060/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0043.png", "metadata/rollout_060/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0044.png", "metadata/rollout_060/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0045.png", "metadata/rollout_060/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0046.png", "metadata/rollout_060/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0047.png", "metadata/rollout_060/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0048.png", "metadata/rollout_060/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0049.png", "metadata/rollout_060/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0050.png", "metadata/rollout_060/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0051.png", "metadata/rollout_060/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0052.png", "metadata/rollout_060/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0053.png", "metadata/rollout_060/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0054.png", "metadata/rollout_060/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_060/agentview/0055.png", "metadata/rollout_060/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0056.png", "metadata/rollout_060/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0057.png", "metadata/rollout_060/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0058.png", "metadata/rollout_060/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_060/agentview/0059.png", "metadata/rollout_060/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_060/agentview/0060.png", "metadata/rollout_060/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0061.png", "metadata/rollout_060/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_060/agentview/0062.png", "metadata/rollout_060/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_060/agentview/0062.png", "metadata/rollout_060/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0000.png", "metadata/rollout_061/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0001.png", "metadata/rollout_061/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0002.png", "metadata/rollout_061/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0003.png", "metadata/rollout_061/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0004.png", "metadata/rollout_061/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0005.png", "metadata/rollout_061/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0006.png", "metadata/rollout_061/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0007.png", "metadata/rollout_061/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0008.png", "metadata/rollout_061/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_061/agentview/0009.png", "metadata/rollout_061/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_061/agentview/0010.png", "metadata/rollout_061/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_061/agentview/0011.png", "metadata/rollout_061/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_061/agentview/0012.png", "metadata/rollout_061/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_061/agentview/0013.png", "metadata/rollout_061/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_061/agentview/0014.png", "metadata/rollout_061/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_061/agentview/0015.png", "metadata/rollout_061/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0016.png", "metadata/rollout_061/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0017.png", "metadata/rollout_061/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0018.png", "metadata/rollout_061/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0019.png", "metadata/rollout_061/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0020.png", "metadata/rollout_061/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0021.png", "metadata/rollout_061/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_061/agentview/0022.png", "metadata/rollout_061/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0023.png", "metadata/rollout_061/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0024.png", "metadata/rollout_061/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0025.png", "metadata/rollout_061/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0026.png", "metadata/rollout_061/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0027.png", "metadata/rollout_061/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0028.png", "metadata/rollout_061/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0029.png", "metadata/rollout_061/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0030.png", "metadata/rollout_061/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_061/agentview/0031.png", "metadata/rollout_061/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_061/agentview/0032.png", "metadata/rollout_061/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_061/agentview/0033.png", "metadata/rollout_061/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_061/agentview/0034.png", "metadata/rollout_061/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_061/agentview/0035.png", "metadata/rollout_061/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_061/agentview/0036.png", "metadata/rollout_061/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_061/agentview/0037.png", "metadata/rollout_061/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_061/agentview/0038.png", "metadata/rollout_061/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_061/agentview/0039.png", "metadata/rollout_061/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_061/agentview/0040.png", "metadata/rollout_061/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0041.png", "metadata/rollout_061/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0042.png", "metadata/rollout_061/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0043.png", "metadata/rollout_061/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_061/agentview/0044.png", "metadata/rollout_061/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_061/agentview/0045.png", "metadata/rollout_061/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0046.png", "metadata/rollout_061/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_061/agentview/0047.png", "metadata/rollout_061/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_061/agentview/0047.png", "metadata/rollout_061/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0000.png", "metadata/rollout_062/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0001.png", "metadata/rollout_062/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0002.png", "metadata/rollout_062/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0003.png", "metadata/rollout_062/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0004.png", "metadata/rollout_062/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0005.png", "metadata/rollout_062/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0006.png", "metadata/rollout_062/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0007.png", "metadata/rollout_062/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0008.png", "metadata/rollout_062/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_062/agentview/0009.png", "metadata/rollout_062/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_062/agentview/0010.png", "metadata/rollout_062/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_062/agentview/0011.png", "metadata/rollout_062/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_062/agentview/0012.png", "metadata/rollout_062/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_062/agentview/0013.png", "metadata/rollout_062/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_062/agentview/0014.png", "metadata/rollout_062/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_062/agentview/0015.png", "metadata/rollout_062/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_062/agentview/0016.png", "metadata/rollout_062/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_062/agentview/0017.png", "metadata/rollout_062/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_062/agentview/0018.png", "metadata/rollout_062/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_062/agentview/0019.png", "metadata/rollout_062/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_062/agentview/0020.png", "metadata/rollout_062/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_062/agentview/0021.png", "metadata/rollout_062/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_062/agentview/0022.png", "metadata/rollout_062/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_062/agentview/0023.png", "metadata/rollout_062/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0024.png", "metadata/rollout_062/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0025.png", "metadata/rollout_062/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0026.png", "metadata/rollout_062/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0027.png", "metadata/rollout_062/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0028.png", "metadata/rollout_062/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_062/agentview/0029.png", "metadata/rollout_062/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_062/agentview/0030.png", "metadata/rollout_062/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_062/agentview/0031.png", "metadata/rollout_062/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_062/agentview/0032.png", "metadata/rollout_062/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_062/agentview/0033.png", "metadata/rollout_062/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_062/agentview/0034.png", "metadata/rollout_062/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_062/agentview/0035.png", "metadata/rollout_062/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_062/agentview/0036.png", "metadata/rollout_062/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_062/agentview/0037.png", "metadata/rollout_062/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_062/agentview/0038.png", "metadata/rollout_062/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_062/agentview/0039.png", "metadata/rollout_062/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0040.png", "metadata/rollout_062/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0041.png", "metadata/rollout_062/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0042.png", "metadata/rollout_062/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0043.png", "metadata/rollout_062/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0044.png", "metadata/rollout_062/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0045.png", "metadata/rollout_062/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0046.png", "metadata/rollout_062/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0047.png", "metadata/rollout_062/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0048.png", "metadata/rollout_062/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0049.png", "metadata/rollout_062/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0050.png", "metadata/rollout_062/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_062/agentview/0051.png", "metadata/rollout_062/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0052.png", "metadata/rollout_062/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0053.png", "metadata/rollout_062/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_062/agentview/0054.png", "metadata/rollout_062/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_062/agentview/0055.png", "metadata/rollout_062/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_062/agentview/0056.png", "metadata/rollout_062/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_062/agentview/0057.png", "metadata/rollout_062/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_062/agentview/0057.png", "metadata/rollout_062/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0000.png", "metadata/rollout_063/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0001.png", "metadata/rollout_063/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0002.png", "metadata/rollout_063/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0003.png", "metadata/rollout_063/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0004.png", "metadata/rollout_063/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0005.png", "metadata/rollout_063/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0006.png", "metadata/rollout_063/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0007.png", "metadata/rollout_063/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0008.png", "metadata/rollout_063/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_063/agentview/0009.png", "metadata/rollout_063/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_063/agentview/0010.png", "metadata/rollout_063/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_063/agentview/0011.png", "metadata/rollout_063/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_063/agentview/0012.png", "metadata/rollout_063/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_063/agentview/0013.png", "metadata/rollout_063/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_063/agentview/0014.png", "metadata/rollout_063/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_063/agentview/0015.png", "metadata/rollout_063/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_063/agentview/0016.png", "metadata/rollout_063/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0017.png", "metadata/rollout_063/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0018.png", "metadata/rollout_063/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0019.png", "metadata/rollout_063/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0020.png", "metadata/rollout_063/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0021.png", "metadata/rollout_063/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0022.png", "metadata/rollout_063/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_063/agentview/0023.png", "metadata/rollout_063/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0024.png", "metadata/rollout_063/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0025.png", "metadata/rollout_063/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0026.png", "metadata/rollout_063/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0027.png", "metadata/rollout_063/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0028.png", "metadata/rollout_063/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0029.png", "metadata/rollout_063/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0030.png", "metadata/rollout_063/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0031.png", "metadata/rollout_063/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_063/agentview/0032.png", "metadata/rollout_063/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_063/agentview/0033.png", "metadata/rollout_063/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_063/agentview/0034.png", "metadata/rollout_063/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_063/agentview/0035.png", "metadata/rollout_063/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_063/agentview/0036.png", "metadata/rollout_063/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_063/agentview/0037.png", "metadata/rollout_063/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_063/agentview/0038.png", "metadata/rollout_063/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_063/agentview/0039.png", "metadata/rollout_063/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_063/agentview/0040.png", "metadata/rollout_063/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_063/agentview/0041.png", "metadata/rollout_063/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0042.png", "metadata/rollout_063/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0043.png", "metadata/rollout_063/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0044.png", "metadata/rollout_063/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_063/agentview/0045.png", "metadata/rollout_063/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_063/agentview/0046.png", "metadata/rollout_063/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0047.png", "metadata/rollout_063/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_063/agentview/0048.png", "metadata/rollout_063/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_063/agentview/0048.png", "metadata/rollout_063/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0000.png", "metadata/rollout_064/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0001.png", "metadata/rollout_064/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0002.png", "metadata/rollout_064/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0003.png", "metadata/rollout_064/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0004.png", "metadata/rollout_064/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0005.png", "metadata/rollout_064/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0006.png", "metadata/rollout_064/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0007.png", "metadata/rollout_064/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0008.png", "metadata/rollout_064/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0009.png", "metadata/rollout_064/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0010.png", "metadata/rollout_064/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0011.png", "metadata/rollout_064/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0012.png", "metadata/rollout_064/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0013.png", "metadata/rollout_064/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0014.png", "metadata/rollout_064/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_064/agentview/0015.png", "metadata/rollout_064/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_064/agentview/0016.png", "metadata/rollout_064/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_064/agentview/0017.png", "metadata/rollout_064/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_064/agentview/0018.png", "metadata/rollout_064/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0019.png", "metadata/rollout_064/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0020.png", "metadata/rollout_064/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0021.png", "metadata/rollout_064/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0022.png", "metadata/rollout_064/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0023.png", "metadata/rollout_064/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0024.png", "metadata/rollout_064/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_064/agentview/0025.png", "metadata/rollout_064/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_064/agentview/0026.png", "metadata/rollout_064/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_064/agentview/0027.png", "metadata/rollout_064/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_064/agentview/0028.png", "metadata/rollout_064/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_064/agentview/0029.png", "metadata/rollout_064/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_064/agentview/0030.png", "metadata/rollout_064/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_064/agentview/0031.png", "metadata/rollout_064/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_064/agentview/0032.png", "metadata/rollout_064/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0033.png", "metadata/rollout_064/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0034.png", "metadata/rollout_064/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0035.png", "metadata/rollout_064/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0036.png", "metadata/rollout_064/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0037.png", "metadata/rollout_064/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0038.png", "metadata/rollout_064/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0039.png", "metadata/rollout_064/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0040.png", "metadata/rollout_064/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0041.png", "metadata/rollout_064/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0042.png", "metadata/rollout_064/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0043.png", "metadata/rollout_064/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0044.png", "metadata/rollout_064/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0045.png", "metadata/rollout_064/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0046.png", "metadata/rollout_064/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_064/agentview/0047.png", "metadata/rollout_064/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_064/agentview/0048.png", "metadata/rollout_064/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0049.png", "metadata/rollout_064/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0050.png", "metadata/rollout_064/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_064/agentview/0051.png", "metadata/rollout_064/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_064/agentview/0052.png", "metadata/rollout_064/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_064/agentview/0053.png", "metadata/rollout_064/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_064/agentview/0054.png", "metadata/rollout_064/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_064/agentview/0054.png", "metadata/rollout_064/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0000.png", "metadata/rollout_065/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0001.png", "metadata/rollout_065/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0002.png", "metadata/rollout_065/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0003.png", "metadata/rollout_065/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0004.png", "metadata/rollout_065/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0005.png", "metadata/rollout_065/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0006.png", "metadata/rollout_065/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0007.png", "metadata/rollout_065/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0008.png", "metadata/rollout_065/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_065/agentview/0009.png", "metadata/rollout_065/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_065/agentview/0010.png", "metadata/rollout_065/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_065/agentview/0011.png", "metadata/rollout_065/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_065/agentview/0012.png", "metadata/rollout_065/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_065/agentview/0013.png", "metadata/rollout_065/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_065/agentview/0014.png", "metadata/rollout_065/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_065/agentview/0015.png", "metadata/rollout_065/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_065/agentview/0016.png", "metadata/rollout_065/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_065/agentview/0017.png", "metadata/rollout_065/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_065/agentview/0018.png", "metadata/rollout_065/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_065/agentview/0019.png", "metadata/rollout_065/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_065/agentview/0020.png", "metadata/rollout_065/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_065/agentview/0021.png", "metadata/rollout_065/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0022.png", "metadata/rollout_065/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0023.png", "metadata/rollout_065/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0024.png", "metadata/rollout_065/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0025.png", "metadata/rollout_065/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0026.png", "metadata/rollout_065/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0027.png", "metadata/rollout_065/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_065/agentview/0028.png", "metadata/rollout_065/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0029.png", "metadata/rollout_065/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0030.png", "metadata/rollout_065/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0031.png", "metadata/rollout_065/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0032.png", "metadata/rollout_065/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0033.png", "metadata/rollout_065/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0034.png", "metadata/rollout_065/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0035.png", "metadata/rollout_065/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0036.png", "metadata/rollout_065/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_065/agentview/0037.png", "metadata/rollout_065/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_065/agentview/0038.png", "metadata/rollout_065/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_065/agentview/0039.png", "metadata/rollout_065/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_065/agentview/0040.png", "metadata/rollout_065/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_065/agentview/0041.png", "metadata/rollout_065/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_065/agentview/0042.png", "metadata/rollout_065/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_065/agentview/0043.png", "metadata/rollout_065/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_065/agentview/0044.png", "metadata/rollout_065/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0045.png", "metadata/rollout_065/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0046.png", "metadata/rollout_065/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0047.png", "metadata/rollout_065/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_065/agentview/0048.png", "metadata/rollout_065/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_065/agentview/0049.png", "metadata/rollout_065/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0050.png", "metadata/rollout_065/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_065/agentview/0051.png", "metadata/rollout_065/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_065/agentview/0051.png", "metadata/rollout_065/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0000.png", "metadata/rollout_066/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0001.png", "metadata/rollout_066/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0002.png", "metadata/rollout_066/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0003.png", "metadata/rollout_066/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0004.png", "metadata/rollout_066/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0005.png", "metadata/rollout_066/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0006.png", "metadata/rollout_066/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0007.png", "metadata/rollout_066/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0008.png", "metadata/rollout_066/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0009.png", "metadata/rollout_066/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0010.png", "metadata/rollout_066/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0011.png", "metadata/rollout_066/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0012.png", "metadata/rollout_066/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0013.png", "metadata/rollout_066/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_066/agentview/0014.png", "metadata/rollout_066/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_066/agentview/0015.png", "metadata/rollout_066/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0016.png", "metadata/rollout_066/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0017.png", "metadata/rollout_066/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0018.png", "metadata/rollout_066/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0019.png", "metadata/rollout_066/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0020.png", "metadata/rollout_066/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_066/agentview/0021.png", "metadata/rollout_066/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_066/agentview/0022.png", "metadata/rollout_066/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_066/agentview/0023.png", "metadata/rollout_066/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_066/agentview/0024.png", "metadata/rollout_066/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_066/agentview/0025.png", "metadata/rollout_066/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_066/agentview/0026.png", "metadata/rollout_066/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_066/agentview/0027.png", "metadata/rollout_066/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_066/agentview/0028.png", "metadata/rollout_066/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_066/agentview/0029.png", "metadata/rollout_066/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_066/agentview/0030.png", "metadata/rollout_066/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_066/agentview/0031.png", "metadata/rollout_066/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_066/agentview/0032.png", "metadata/rollout_066/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_066/agentview/0033.png", "metadata/rollout_066/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_066/agentview/0034.png", "metadata/rollout_066/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0035.png", "metadata/rollout_066/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0036.png", "metadata/rollout_066/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0037.png", "metadata/rollout_066/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0038.png", "metadata/rollout_066/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0039.png", "metadata/rollout_066/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_066/agentview/0040.png", "metadata/rollout_066/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0041.png", "metadata/rollout_066/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0042.png", "metadata/rollout_066/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_066/agentview/0043.png", "metadata/rollout_066/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_066/agentview/0044.png", "metadata/rollout_066/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_066/agentview/0045.png", "metadata/rollout_066/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_066/agentview/0046.png", "metadata/rollout_066/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_066/agentview/0046.png", "metadata/rollout_066/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0000.png", "metadata/rollout_067/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0001.png", "metadata/rollout_067/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0002.png", "metadata/rollout_067/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0003.png", "metadata/rollout_067/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0004.png", "metadata/rollout_067/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0005.png", "metadata/rollout_067/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0006.png", "metadata/rollout_067/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0007.png", "metadata/rollout_067/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0008.png", "metadata/rollout_067/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0009.png", "metadata/rollout_067/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0010.png", "metadata/rollout_067/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0011.png", "metadata/rollout_067/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0012.png", "metadata/rollout_067/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0013.png", "metadata/rollout_067/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0014.png", "metadata/rollout_067/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0015.png", "metadata/rollout_067/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_067/agentview/0016.png", "metadata/rollout_067/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0017.png", "metadata/rollout_067/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0018.png", "metadata/rollout_067/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0019.png", "metadata/rollout_067/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0020.png", "metadata/rollout_067/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0021.png", "metadata/rollout_067/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0022.png", "metadata/rollout_067/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_067/agentview/0023.png", "metadata/rollout_067/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0024.png", "metadata/rollout_067/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0025.png", "metadata/rollout_067/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0026.png", "metadata/rollout_067/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0027.png", "metadata/rollout_067/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0028.png", "metadata/rollout_067/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0029.png", "metadata/rollout_067/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0030.png", "metadata/rollout_067/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0031.png", "metadata/rollout_067/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0032.png", "metadata/rollout_067/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0033.png", "metadata/rollout_067/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_067/agentview/0034.png", "metadata/rollout_067/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_067/agentview/0035.png", "metadata/rollout_067/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_067/agentview/0036.png", "metadata/rollout_067/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_067/agentview/0037.png", "metadata/rollout_067/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_067/agentview/0038.png", "metadata/rollout_067/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_067/agentview/0039.png", "metadata/rollout_067/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_067/agentview/0040.png", "metadata/rollout_067/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_067/agentview/0041.png", "metadata/rollout_067/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0042.png", "metadata/rollout_067/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0043.png", "metadata/rollout_067/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0044.png", "metadata/rollout_067/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_067/agentview/0045.png", "metadata/rollout_067/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_067/agentview/0046.png", "metadata/rollout_067/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0047.png", "metadata/rollout_067/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_067/agentview/0048.png", "metadata/rollout_067/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_067/agentview/0048.png", "metadata/rollout_067/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0000.png", "metadata/rollout_068/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0001.png", "metadata/rollout_068/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0002.png", "metadata/rollout_068/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0003.png", "metadata/rollout_068/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0004.png", "metadata/rollout_068/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0005.png", "metadata/rollout_068/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0006.png", "metadata/rollout_068/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0007.png", "metadata/rollout_068/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0008.png", "metadata/rollout_068/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_068/agentview/0009.png", "metadata/rollout_068/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_068/agentview/0010.png", "metadata/rollout_068/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_068/agentview/0011.png", "metadata/rollout_068/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_068/agentview/0012.png", "metadata/rollout_068/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_068/agentview/0013.png", "metadata/rollout_068/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_068/agentview/0014.png", "metadata/rollout_068/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_068/agentview/0015.png", "metadata/rollout_068/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_068/agentview/0016.png", "metadata/rollout_068/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0017.png", "metadata/rollout_068/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0018.png", "metadata/rollout_068/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0019.png", "metadata/rollout_068/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0020.png", "metadata/rollout_068/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0021.png", "metadata/rollout_068/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0022.png", "metadata/rollout_068/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_068/agentview/0023.png", "metadata/rollout_068/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0024.png", "metadata/rollout_068/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0025.png", "metadata/rollout_068/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0026.png", "metadata/rollout_068/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0027.png", "metadata/rollout_068/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0028.png", "metadata/rollout_068/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0029.png", "metadata/rollout_068/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0030.png", "metadata/rollout_068/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0031.png", "metadata/rollout_068/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_068/agentview/0032.png", "metadata/rollout_068/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_068/agentview/0033.png", "metadata/rollout_068/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_068/agentview/0034.png", "metadata/rollout_068/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_068/agentview/0035.png", "metadata/rollout_068/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0036.png", "metadata/rollout_068/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0037.png", "metadata/rollout_068/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0038.png", "metadata/rollout_068/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0039.png", "metadata/rollout_068/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0040.png", "metadata/rollout_068/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0041.png", "metadata/rollout_068/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0042.png", "metadata/rollout_068/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0043.png", "metadata/rollout_068/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0044.png", "metadata/rollout_068/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_068/agentview/0045.png", "metadata/rollout_068/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_068/agentview/0046.png", "metadata/rollout_068/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0047.png", "metadata/rollout_068/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0048.png", "metadata/rollout_068/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0049.png", "metadata/rollout_068/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_068/agentview/0050.png", "metadata/rollout_068/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_068/agentview/0051.png", "metadata/rollout_068/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0052.png", "metadata/rollout_068/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_068/agentview/0053.png", "metadata/rollout_068/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_068/agentview/0053.png", "metadata/rollout_068/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0000.png", "metadata/rollout_069/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0001.png", "metadata/rollout_069/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0002.png", "metadata/rollout_069/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0003.png", "metadata/rollout_069/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0004.png", "metadata/rollout_069/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0005.png", "metadata/rollout_069/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0006.png", "metadata/rollout_069/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0007.png", "metadata/rollout_069/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0008.png", "metadata/rollout_069/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_069/agentview/0009.png", "metadata/rollout_069/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_069/agentview/0010.png", "metadata/rollout_069/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_069/agentview/0011.png", "metadata/rollout_069/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_069/agentview/0012.png", "metadata/rollout_069/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_069/agentview/0013.png", "metadata/rollout_069/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_069/agentview/0014.png", "metadata/rollout_069/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_069/agentview/0015.png", "metadata/rollout_069/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0016.png", "metadata/rollout_069/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0017.png", "metadata/rollout_069/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0018.png", "metadata/rollout_069/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0019.png", "metadata/rollout_069/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0020.png", "metadata/rollout_069/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_069/agentview/0021.png", "metadata/rollout_069/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_069/agentview/0022.png", "metadata/rollout_069/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_069/agentview/0023.png", "metadata/rollout_069/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_069/agentview/0024.png", "metadata/rollout_069/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_069/agentview/0025.png", "metadata/rollout_069/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_069/agentview/0026.png", "metadata/rollout_069/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_069/agentview/0027.png", "metadata/rollout_069/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_069/agentview/0028.png", "metadata/rollout_069/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_069/agentview/0029.png", "metadata/rollout_069/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_069/agentview/0030.png", "metadata/rollout_069/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_069/agentview/0031.png", "metadata/rollout_069/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_069/agentview/0032.png", "metadata/rollout_069/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_069/agentview/0033.png", "metadata/rollout_069/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_069/agentview/0034.png", "metadata/rollout_069/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_069/agentview/0035.png", "metadata/rollout_069/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_069/agentview/0036.png", "metadata/rollout_069/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_069/agentview/0037.png", "metadata/rollout_069/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_069/agentview/0038.png", "metadata/rollout_069/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0039.png", "metadata/rollout_069/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0040.png", "metadata/rollout_069/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_069/agentview/0041.png", "metadata/rollout_069/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_069/agentview/0042.png", "metadata/rollout_069/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_069/agentview/0043.png", "metadata/rollout_069/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_069/agentview/0044.png", "metadata/rollout_069/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the orange block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_069/agentview/0044.png", "metadata/rollout_069/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0000.png", "metadata/rollout_070/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0001.png", "metadata/rollout_070/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0002.png", "metadata/rollout_070/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0003.png", "metadata/rollout_070/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0004.png", "metadata/rollout_070/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0005.png", "metadata/rollout_070/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0006.png", "metadata/rollout_070/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0007.png", "metadata/rollout_070/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0008.png", "metadata/rollout_070/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_070/agentview/0009.png", "metadata/rollout_070/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_070/agentview/0010.png", "metadata/rollout_070/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_070/agentview/0011.png", "metadata/rollout_070/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_070/agentview/0012.png", "metadata/rollout_070/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_070/agentview/0013.png", "metadata/rollout_070/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_070/agentview/0014.png", "metadata/rollout_070/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_070/agentview/0015.png", "metadata/rollout_070/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0016.png", "metadata/rollout_070/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0017.png", "metadata/rollout_070/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0018.png", "metadata/rollout_070/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0019.png", "metadata/rollout_070/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0020.png", "metadata/rollout_070/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0021.png", "metadata/rollout_070/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_070/agentview/0022.png", "metadata/rollout_070/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0023.png", "metadata/rollout_070/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0024.png", "metadata/rollout_070/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0025.png", "metadata/rollout_070/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0026.png", "metadata/rollout_070/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0027.png", "metadata/rollout_070/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0028.png", "metadata/rollout_070/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0029.png", "metadata/rollout_070/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0030.png", "metadata/rollout_070/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_070/agentview/0031.png", "metadata/rollout_070/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_070/agentview/0032.png", "metadata/rollout_070/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_070/agentview/0033.png", "metadata/rollout_070/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_070/agentview/0034.png", "metadata/rollout_070/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_070/agentview/0035.png", "metadata/rollout_070/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_070/agentview/0036.png", "metadata/rollout_070/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_070/agentview/0037.png", "metadata/rollout_070/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_070/agentview/0038.png", "metadata/rollout_070/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_070/agentview/0039.png", "metadata/rollout_070/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0040.png", "metadata/rollout_070/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0041.png", "metadata/rollout_070/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0042.png", "metadata/rollout_070/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_070/agentview/0043.png", "metadata/rollout_070/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_070/agentview/0044.png", "metadata/rollout_070/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0045.png", "metadata/rollout_070/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_070/agentview/0046.png", "metadata/rollout_070/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_070/agentview/0046.png", "metadata/rollout_070/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0000.png", "metadata/rollout_071/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0001.png", "metadata/rollout_071/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0002.png", "metadata/rollout_071/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0003.png", "metadata/rollout_071/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0004.png", "metadata/rollout_071/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0005.png", "metadata/rollout_071/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0006.png", "metadata/rollout_071/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0007.png", "metadata/rollout_071/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0008.png", "metadata/rollout_071/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_071/agentview/0009.png", "metadata/rollout_071/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_071/agentview/0010.png", "metadata/rollout_071/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_071/agentview/0011.png", "metadata/rollout_071/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_071/agentview/0012.png", "metadata/rollout_071/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_071/agentview/0013.png", "metadata/rollout_071/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_071/agentview/0014.png", "metadata/rollout_071/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_071/agentview/0015.png", "metadata/rollout_071/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_071/agentview/0016.png", "metadata/rollout_071/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_071/agentview/0017.png", "metadata/rollout_071/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0018.png", "metadata/rollout_071/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0019.png", "metadata/rollout_071/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0020.png", "metadata/rollout_071/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0021.png", "metadata/rollout_071/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0022.png", "metadata/rollout_071/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0023.png", "metadata/rollout_071/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_071/agentview/0024.png", "metadata/rollout_071/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0025.png", "metadata/rollout_071/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0026.png", "metadata/rollout_071/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0027.png", "metadata/rollout_071/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0028.png", "metadata/rollout_071/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0029.png", "metadata/rollout_071/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0030.png", "metadata/rollout_071/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0031.png", "metadata/rollout_071/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0032.png", "metadata/rollout_071/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0033.png", "metadata/rollout_071/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0034.png", "metadata/rollout_071/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0035.png", "metadata/rollout_071/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0036.png", "metadata/rollout_071/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0037.png", "metadata/rollout_071/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0038.png", "metadata/rollout_071/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0039.png", "metadata/rollout_071/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0040.png", "metadata/rollout_071/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0041.png", "metadata/rollout_071/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0042.png", "metadata/rollout_071/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_071/agentview/0043.png", "metadata/rollout_071/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_071/agentview/0044.png", "metadata/rollout_071/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_071/agentview/0045.png", "metadata/rollout_071/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_071/agentview/0046.png", "metadata/rollout_071/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0047.png", "metadata/rollout_071/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0048.png", "metadata/rollout_071/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0049.png", "metadata/rollout_071/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_071/agentview/0050.png", "metadata/rollout_071/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_071/agentview/0051.png", "metadata/rollout_071/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0052.png", "metadata/rollout_071/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_071/agentview/0053.png", "metadata/rollout_071/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_071/agentview/0053.png", "metadata/rollout_071/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0000.png", "metadata/rollout_072/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0001.png", "metadata/rollout_072/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0002.png", "metadata/rollout_072/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0003.png", "metadata/rollout_072/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0004.png", "metadata/rollout_072/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0005.png", "metadata/rollout_072/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0006.png", "metadata/rollout_072/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0007.png", "metadata/rollout_072/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0008.png", "metadata/rollout_072/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_072/agentview/0009.png", "metadata/rollout_072/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_072/agentview/0010.png", "metadata/rollout_072/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_072/agentview/0011.png", "metadata/rollout_072/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_072/agentview/0012.png", "metadata/rollout_072/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_072/agentview/0013.png", "metadata/rollout_072/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_072/agentview/0014.png", "metadata/rollout_072/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_072/agentview/0015.png", "metadata/rollout_072/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_072/agentview/0016.png", "metadata/rollout_072/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_072/agentview/0017.png", "metadata/rollout_072/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_072/agentview/0018.png", "metadata/rollout_072/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_072/agentview/0019.png", "metadata/rollout_072/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_072/agentview/0020.png", "metadata/rollout_072/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_072/agentview/0021.png", "metadata/rollout_072/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_072/agentview/0022.png", "metadata/rollout_072/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_072/agentview/0023.png", "metadata/rollout_072/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_072/agentview/0024.png", "metadata/rollout_072/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_072/agentview/0025.png", "metadata/rollout_072/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0026.png", "metadata/rollout_072/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0027.png", "metadata/rollout_072/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0028.png", "metadata/rollout_072/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0029.png", "metadata/rollout_072/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0030.png", "metadata/rollout_072/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0031.png", "metadata/rollout_072/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_072/agentview/0032.png", "metadata/rollout_072/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0033.png", "metadata/rollout_072/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0034.png", "metadata/rollout_072/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0035.png", "metadata/rollout_072/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0036.png", "metadata/rollout_072/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0037.png", "metadata/rollout_072/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0038.png", "metadata/rollout_072/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0039.png", "metadata/rollout_072/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0040.png", "metadata/rollout_072/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_072/agentview/0041.png", "metadata/rollout_072/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_072/agentview/0042.png", "metadata/rollout_072/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_072/agentview/0043.png", "metadata/rollout_072/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0044.png", "metadata/rollout_072/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0045.png", "metadata/rollout_072/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0046.png", "metadata/rollout_072/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0047.png", "metadata/rollout_072/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0048.png", "metadata/rollout_072/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0049.png", "metadata/rollout_072/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0050.png", "metadata/rollout_072/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0051.png", "metadata/rollout_072/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0052.png", "metadata/rollout_072/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0053.png", "metadata/rollout_072/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0054.png", "metadata/rollout_072/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0055.png", "metadata/rollout_072/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_072/agentview/0056.png", "metadata/rollout_072/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0057.png", "metadata/rollout_072/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0058.png", "metadata/rollout_072/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0059.png", "metadata/rollout_072/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_072/agentview/0060.png", "metadata/rollout_072/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_072/agentview/0061.png", "metadata/rollout_072/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0062.png", "metadata/rollout_072/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_072/agentview/0063.png", "metadata/rollout_072/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_072/agentview/0063.png", "metadata/rollout_072/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0000.png", "metadata/rollout_073/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0001.png", "metadata/rollout_073/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0002.png", "metadata/rollout_073/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0003.png", "metadata/rollout_073/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0004.png", "metadata/rollout_073/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0005.png", "metadata/rollout_073/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0006.png", "metadata/rollout_073/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0007.png", "metadata/rollout_073/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0008.png", "metadata/rollout_073/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0009.png", "metadata/rollout_073/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0010.png", "metadata/rollout_073/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0011.png", "metadata/rollout_073/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0012.png", "metadata/rollout_073/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0013.png", "metadata/rollout_073/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0014.png", "metadata/rollout_073/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0015.png", "metadata/rollout_073/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0016.png", "metadata/rollout_073/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0017.png", "metadata/rollout_073/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_073/agentview/0018.png", "metadata/rollout_073/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_073/agentview/0019.png", "metadata/rollout_073/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_073/agentview/0020.png", "metadata/rollout_073/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_073/agentview/0021.png", "metadata/rollout_073/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_073/agentview/0022.png", "metadata/rollout_073/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_073/agentview/0023.png", "metadata/rollout_073/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_073/agentview/0024.png", "metadata/rollout_073/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0025.png", "metadata/rollout_073/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0026.png", "metadata/rollout_073/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0027.png", "metadata/rollout_073/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0028.png", "metadata/rollout_073/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0029.png", "metadata/rollout_073/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0030.png", "metadata/rollout_073/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_073/agentview/0031.png", "metadata/rollout_073/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0032.png", "metadata/rollout_073/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0033.png", "metadata/rollout_073/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0034.png", "metadata/rollout_073/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0035.png", "metadata/rollout_073/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0036.png", "metadata/rollout_073/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0037.png", "metadata/rollout_073/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0038.png", "metadata/rollout_073/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0039.png", "metadata/rollout_073/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0040.png", "metadata/rollout_073/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0041.png", "metadata/rollout_073/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0042.png", "metadata/rollout_073/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0043.png", "metadata/rollout_073/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0044.png", "metadata/rollout_073/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0045.png", "metadata/rollout_073/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0046.png", "metadata/rollout_073/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0047.png", "metadata/rollout_073/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0048.png", "metadata/rollout_073/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0049.png", "metadata/rollout_073/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0050.png", "metadata/rollout_073/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0051.png", "metadata/rollout_073/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0052.png", "metadata/rollout_073/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_073/agentview/0053.png", "metadata/rollout_073/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_073/agentview/0054.png", "metadata/rollout_073/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0055.png", "metadata/rollout_073/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0056.png", "metadata/rollout_073/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0057.png", "metadata/rollout_073/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_073/agentview/0058.png", "metadata/rollout_073/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_073/agentview/0059.png", "metadata/rollout_073/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0060.png", "metadata/rollout_073/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_073/agentview/0061.png", "metadata/rollout_073/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_073/agentview/0061.png", "metadata/rollout_073/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0000.png", "metadata/rollout_074/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0001.png", "metadata/rollout_074/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0002.png", "metadata/rollout_074/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0003.png", "metadata/rollout_074/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0004.png", "metadata/rollout_074/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0005.png", "metadata/rollout_074/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0006.png", "metadata/rollout_074/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0007.png", "metadata/rollout_074/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0008.png", "metadata/rollout_074/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_074/agentview/0009.png", "metadata/rollout_074/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_074/agentview/0010.png", "metadata/rollout_074/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_074/agentview/0011.png", "metadata/rollout_074/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_074/agentview/0012.png", "metadata/rollout_074/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_074/agentview/0013.png", "metadata/rollout_074/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_074/agentview/0014.png", "metadata/rollout_074/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_074/agentview/0015.png", "metadata/rollout_074/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_074/agentview/0016.png", "metadata/rollout_074/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_074/agentview/0017.png", "metadata/rollout_074/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_074/agentview/0018.png", "metadata/rollout_074/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_074/agentview/0019.png", "metadata/rollout_074/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_074/agentview/0020.png", "metadata/rollout_074/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0021.png", "metadata/rollout_074/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0022.png", "metadata/rollout_074/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0023.png", "metadata/rollout_074/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0024.png", "metadata/rollout_074/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0025.png", "metadata/rollout_074/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0026.png", "metadata/rollout_074/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_074/agentview/0027.png", "metadata/rollout_074/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0028.png", "metadata/rollout_074/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0029.png", "metadata/rollout_074/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0030.png", "metadata/rollout_074/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0031.png", "metadata/rollout_074/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0032.png", "metadata/rollout_074/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0033.png", "metadata/rollout_074/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0034.png", "metadata/rollout_074/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0035.png", "metadata/rollout_074/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_074/agentview/0036.png", "metadata/rollout_074/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_074/agentview/0037.png", "metadata/rollout_074/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_074/agentview/0038.png", "metadata/rollout_074/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_074/agentview/0039.png", "metadata/rollout_074/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_074/agentview/0040.png", "metadata/rollout_074/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_074/agentview/0041.png", "metadata/rollout_074/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_074/agentview/0042.png", "metadata/rollout_074/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_074/agentview/0043.png", "metadata/rollout_074/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_074/agentview/0044.png", "metadata/rollout_074/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_074/agentview/0045.png", "metadata/rollout_074/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0046.png", "metadata/rollout_074/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0047.png", "metadata/rollout_074/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0048.png", "metadata/rollout_074/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_074/agentview/0049.png", "metadata/rollout_074/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_074/agentview/0050.png", "metadata/rollout_074/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0051.png", "metadata/rollout_074/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_074/agentview/0052.png", "metadata/rollout_074/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_074/agentview/0052.png", "metadata/rollout_074/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0000.png", "metadata/rollout_075/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0001.png", "metadata/rollout_075/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0002.png", "metadata/rollout_075/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0003.png", "metadata/rollout_075/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0004.png", "metadata/rollout_075/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0005.png", "metadata/rollout_075/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0006.png", "metadata/rollout_075/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0007.png", "metadata/rollout_075/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0008.png", "metadata/rollout_075/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_075/agentview/0009.png", "metadata/rollout_075/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_075/agentview/0010.png", "metadata/rollout_075/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_075/agentview/0011.png", "metadata/rollout_075/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_075/agentview/0012.png", "metadata/rollout_075/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_075/agentview/0013.png", "metadata/rollout_075/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0014.png", "metadata/rollout_075/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0015.png", "metadata/rollout_075/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0016.png", "metadata/rollout_075/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0017.png", "metadata/rollout_075/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0018.png", "metadata/rollout_075/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0019.png", "metadata/rollout_075/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_075/agentview/0020.png", "metadata/rollout_075/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0021.png", "metadata/rollout_075/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0022.png", "metadata/rollout_075/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0023.png", "metadata/rollout_075/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0024.png", "metadata/rollout_075/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0025.png", "metadata/rollout_075/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0026.png", "metadata/rollout_075/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0027.png", "metadata/rollout_075/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0028.png", "metadata/rollout_075/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0029.png", "metadata/rollout_075/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0030.png", "metadata/rollout_075/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0031.png", "metadata/rollout_075/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0032.png", "metadata/rollout_075/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0033.png", "metadata/rollout_075/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0034.png", "metadata/rollout_075/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0035.png", "metadata/rollout_075/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0036.png", "metadata/rollout_075/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0037.png", "metadata/rollout_075/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0038.png", "metadata/rollout_075/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_075/agentview/0039.png", "metadata/rollout_075/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0040.png", "metadata/rollout_075/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0041.png", "metadata/rollout_075/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0042.png", "metadata/rollout_075/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_075/agentview/0043.png", "metadata/rollout_075/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_075/agentview/0044.png", "metadata/rollout_075/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0045.png", "metadata/rollout_075/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_075/agentview/0046.png", "metadata/rollout_075/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_075/agentview/0046.png", "metadata/rollout_075/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0000.png", "metadata/rollout_076/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0001.png", "metadata/rollout_076/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0002.png", "metadata/rollout_076/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0003.png", "metadata/rollout_076/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0004.png", "metadata/rollout_076/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0005.png", "metadata/rollout_076/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0006.png", "metadata/rollout_076/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0007.png", "metadata/rollout_076/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0008.png", "metadata/rollout_076/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0009.png", "metadata/rollout_076/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_076/agentview/0010.png", "metadata/rollout_076/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0011.png", "metadata/rollout_076/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0012.png", "metadata/rollout_076/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0013.png", "metadata/rollout_076/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0014.png", "metadata/rollout_076/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0015.png", "metadata/rollout_076/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0016.png", "metadata/rollout_076/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0017.png", "metadata/rollout_076/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0018.png", "metadata/rollout_076/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0019.png", "metadata/rollout_076/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_076/agentview/0020.png", "metadata/rollout_076/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_076/agentview/0021.png", "metadata/rollout_076/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_076/agentview/0022.png", "metadata/rollout_076/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_076/agentview/0023.png", "metadata/rollout_076/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_076/agentview/0024.png", "metadata/rollout_076/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_076/agentview/0025.png", "metadata/rollout_076/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0026.png", "metadata/rollout_076/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0027.png", "metadata/rollout_076/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0028.png", "metadata/rollout_076/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0029.png", "metadata/rollout_076/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0030.png", "metadata/rollout_076/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_076/agentview/0031.png", "metadata/rollout_076/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_076/agentview/0032.png", "metadata/rollout_076/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_076/agentview/0033.png", "metadata/rollout_076/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_076/agentview/0034.png", "metadata/rollout_076/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_076/agentview/0035.png", "metadata/rollout_076/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_076/agentview/0036.png", "metadata/rollout_076/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_076/agentview/0037.png", "metadata/rollout_076/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_076/agentview/0038.png", "metadata/rollout_076/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0039.png", "metadata/rollout_076/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0040.png", "metadata/rollout_076/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0041.png", "metadata/rollout_076/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0042.png", "metadata/rollout_076/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0043.png", "metadata/rollout_076/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0044.png", "metadata/rollout_076/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0045.png", "metadata/rollout_076/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0046.png", "metadata/rollout_076/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0047.png", "metadata/rollout_076/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0048.png", "metadata/rollout_076/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0049.png", "metadata/rollout_076/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0050.png", "metadata/rollout_076/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0051.png", "metadata/rollout_076/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0052.png", "metadata/rollout_076/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_076/agentview/0053.png", "metadata/rollout_076/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_076/agentview/0054.png", "metadata/rollout_076/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_076/agentview/0055.png", "metadata/rollout_076/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_076/agentview/0056.png", "metadata/rollout_076/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_076/agentview/0057.png", "metadata/rollout_076/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_076/agentview/0058.png", "metadata/rollout_076/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0059.png", "metadata/rollout_076/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0060.png", "metadata/rollout_076/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_076/agentview/0061.png", "metadata/rollout_076/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_076/agentview/0062.png", "metadata/rollout_076/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_076/agentview/0063.png", "metadata/rollout_076/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_076/agentview/0064.png", "metadata/rollout_076/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_076/agentview/0064.png", "metadata/rollout_076/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0000.png", "metadata/rollout_077/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0001.png", "metadata/rollout_077/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0002.png", "metadata/rollout_077/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0003.png", "metadata/rollout_077/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0004.png", "metadata/rollout_077/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0005.png", "metadata/rollout_077/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0006.png", "metadata/rollout_077/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0007.png", "metadata/rollout_077/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0008.png", "metadata/rollout_077/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_077/agentview/0009.png", "metadata/rollout_077/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0010.png", "metadata/rollout_077/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0011.png", "metadata/rollout_077/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0012.png", "metadata/rollout_077/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0013.png", "metadata/rollout_077/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0014.png", "metadata/rollout_077/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0015.png", "metadata/rollout_077/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0016.png", "metadata/rollout_077/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0017.png", "metadata/rollout_077/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0018.png", "metadata/rollout_077/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_077/agentview/0019.png", "metadata/rollout_077/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_077/agentview/0020.png", "metadata/rollout_077/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_077/agentview/0021.png", "metadata/rollout_077/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_077/agentview/0022.png", "metadata/rollout_077/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_077/agentview/0023.png", "metadata/rollout_077/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_077/agentview/0024.png", "metadata/rollout_077/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_077/agentview/0025.png", "metadata/rollout_077/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_077/agentview/0026.png", "metadata/rollout_077/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_077/agentview/0027.png", "metadata/rollout_077/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0028.png", "metadata/rollout_077/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0029.png", "metadata/rollout_077/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0030.png", "metadata/rollout_077/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0031.png", "metadata/rollout_077/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0032.png", "metadata/rollout_077/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0033.png", "metadata/rollout_077/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_077/agentview/0034.png", "metadata/rollout_077/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0035.png", "metadata/rollout_077/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0036.png", "metadata/rollout_077/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0037.png", "metadata/rollout_077/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0038.png", "metadata/rollout_077/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0039.png", "metadata/rollout_077/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0040.png", "metadata/rollout_077/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0041.png", "metadata/rollout_077/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0042.png", "metadata/rollout_077/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0043.png", "metadata/rollout_077/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0044.png", "metadata/rollout_077/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0045.png", "metadata/rollout_077/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0046.png", "metadata/rollout_077/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0047.png", "metadata/rollout_077/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0048.png", "metadata/rollout_077/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0049.png", "metadata/rollout_077/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0050.png", "metadata/rollout_077/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0051.png", "metadata/rollout_077/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0052.png", "metadata/rollout_077/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0053.png", "metadata/rollout_077/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0054.png", "metadata/rollout_077/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0055.png", "metadata/rollout_077/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0056.png", "metadata/rollout_077/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0057.png", "metadata/rollout_077/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_077/agentview/0058.png", "metadata/rollout_077/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0059.png", "metadata/rollout_077/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0060.png", "metadata/rollout_077/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0061.png", "metadata/rollout_077/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_077/agentview/0062.png", "metadata/rollout_077/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_077/agentview/0063.png", "metadata/rollout_077/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0064.png", "metadata/rollout_077/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_077/agentview/0065.png", "metadata/rollout_077/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_077/agentview/0065.png", "metadata/rollout_077/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0000.png", "metadata/rollout_078/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0001.png", "metadata/rollout_078/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0002.png", "metadata/rollout_078/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0003.png", "metadata/rollout_078/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0004.png", "metadata/rollout_078/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0005.png", "metadata/rollout_078/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0006.png", "metadata/rollout_078/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0007.png", "metadata/rollout_078/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0008.png", "metadata/rollout_078/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_078/agentview/0009.png", "metadata/rollout_078/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0010.png", "metadata/rollout_078/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0011.png", "metadata/rollout_078/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0012.png", "metadata/rollout_078/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0013.png", "metadata/rollout_078/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0014.png", "metadata/rollout_078/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0015.png", "metadata/rollout_078/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0016.png", "metadata/rollout_078/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0017.png", "metadata/rollout_078/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0018.png", "metadata/rollout_078/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_078/agentview/0019.png", "metadata/rollout_078/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_078/agentview/0020.png", "metadata/rollout_078/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_078/agentview/0021.png", "metadata/rollout_078/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0022.png", "metadata/rollout_078/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0023.png", "metadata/rollout_078/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0024.png", "metadata/rollout_078/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0025.png", "metadata/rollout_078/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0026.png", "metadata/rollout_078/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0027.png", "metadata/rollout_078/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_078/agentview/0028.png", "metadata/rollout_078/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0029.png", "metadata/rollout_078/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0030.png", "metadata/rollout_078/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0031.png", "metadata/rollout_078/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0032.png", "metadata/rollout_078/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0033.png", "metadata/rollout_078/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0034.png", "metadata/rollout_078/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0035.png", "metadata/rollout_078/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0036.png", "metadata/rollout_078/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0037.png", "metadata/rollout_078/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0038.png", "metadata/rollout_078/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0039.png", "metadata/rollout_078/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0040.png", "metadata/rollout_078/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0041.png", "metadata/rollout_078/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0042.png", "metadata/rollout_078/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0043.png", "metadata/rollout_078/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0044.png", "metadata/rollout_078/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0045.png", "metadata/rollout_078/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0046.png", "metadata/rollout_078/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_078/agentview/0047.png", "metadata/rollout_078/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0048.png", "metadata/rollout_078/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0049.png", "metadata/rollout_078/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0050.png", "metadata/rollout_078/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_078/agentview/0051.png", "metadata/rollout_078/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_078/agentview/0052.png", "metadata/rollout_078/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0053.png", "metadata/rollout_078/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_078/agentview/0054.png", "metadata/rollout_078/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_078/agentview/0054.png", "metadata/rollout_078/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0000.png", "metadata/rollout_079/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0001.png", "metadata/rollout_079/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0002.png", "metadata/rollout_079/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0003.png", "metadata/rollout_079/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0004.png", "metadata/rollout_079/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0005.png", "metadata/rollout_079/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0006.png", "metadata/rollout_079/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0007.png", "metadata/rollout_079/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0008.png", "metadata/rollout_079/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_079/agentview/0009.png", "metadata/rollout_079/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0010.png", "metadata/rollout_079/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0011.png", "metadata/rollout_079/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0012.png", "metadata/rollout_079/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0013.png", "metadata/rollout_079/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0014.png", "metadata/rollout_079/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0015.png", "metadata/rollout_079/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0016.png", "metadata/rollout_079/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0017.png", "metadata/rollout_079/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0018.png", "metadata/rollout_079/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_079/agentview/0019.png", "metadata/rollout_079/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_079/agentview/0020.png", "metadata/rollout_079/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_079/agentview/0021.png", "metadata/rollout_079/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_079/agentview/0022.png", "metadata/rollout_079/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_079/agentview/0023.png", "metadata/rollout_079/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_079/agentview/0024.png", "metadata/rollout_079/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0025.png", "metadata/rollout_079/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0026.png", "metadata/rollout_079/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0027.png", "metadata/rollout_079/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0028.png", "metadata/rollout_079/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0029.png", "metadata/rollout_079/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0030.png", "metadata/rollout_079/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_079/agentview/0031.png", "metadata/rollout_079/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0032.png", "metadata/rollout_079/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0033.png", "metadata/rollout_079/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0034.png", "metadata/rollout_079/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0035.png", "metadata/rollout_079/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0036.png", "metadata/rollout_079/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0037.png", "metadata/rollout_079/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0038.png", "metadata/rollout_079/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0039.png", "metadata/rollout_079/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0040.png", "metadata/rollout_079/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0041.png", "metadata/rollout_079/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0042.png", "metadata/rollout_079/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0043.png", "metadata/rollout_079/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0044.png", "metadata/rollout_079/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0045.png", "metadata/rollout_079/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0046.png", "metadata/rollout_079/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0047.png", "metadata/rollout_079/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0048.png", "metadata/rollout_079/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0049.png", "metadata/rollout_079/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_079/agentview/0050.png", "metadata/rollout_079/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_079/agentview/0051.png", "metadata/rollout_079/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0052.png", "metadata/rollout_079/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0053.png", "metadata/rollout_079/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0054.png", "metadata/rollout_079/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_079/agentview/0055.png", "metadata/rollout_079/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_079/agentview/0056.png", "metadata/rollout_079/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0057.png", "metadata/rollout_079/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_079/agentview/0058.png", "metadata/rollout_079/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the red block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_079/agentview/0058.png", "metadata/rollout_079/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0000.png", "metadata/rollout_080/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0001.png", "metadata/rollout_080/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0002.png", "metadata/rollout_080/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0003.png", "metadata/rollout_080/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0004.png", "metadata/rollout_080/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0005.png", "metadata/rollout_080/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0006.png", "metadata/rollout_080/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0007.png", "metadata/rollout_080/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0008.png", "metadata/rollout_080/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0009.png", "metadata/rollout_080/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0010.png", "metadata/rollout_080/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0011.png", "metadata/rollout_080/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0012.png", "metadata/rollout_080/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_080/agentview/0013.png", "metadata/rollout_080/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0014.png", "metadata/rollout_080/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0015.png", "metadata/rollout_080/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0016.png", "metadata/rollout_080/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0017.png", "metadata/rollout_080/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0018.png", "metadata/rollout_080/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_080/agentview/0019.png", "metadata/rollout_080/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_080/agentview/0020.png", "metadata/rollout_080/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_080/agentview/0021.png", "metadata/rollout_080/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_080/agentview/0022.png", "metadata/rollout_080/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_080/agentview/0023.png", "metadata/rollout_080/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_080/agentview/0024.png", "metadata/rollout_080/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_080/agentview/0025.png", "metadata/rollout_080/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_080/agentview/0026.png", "metadata/rollout_080/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_080/agentview/0027.png", "metadata/rollout_080/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_080/agentview/0028.png", "metadata/rollout_080/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_080/agentview/0029.png", "metadata/rollout_080/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_080/agentview/0030.png", "metadata/rollout_080/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0031.png", "metadata/rollout_080/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0032.png", "metadata/rollout_080/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0033.png", "metadata/rollout_080/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0034.png", "metadata/rollout_080/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0035.png", "metadata/rollout_080/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_080/agentview/0036.png", "metadata/rollout_080/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0037.png", "metadata/rollout_080/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_080/agentview/0038.png", "metadata/rollout_080/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_080/agentview/0039.png", "metadata/rollout_080/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_080/agentview/0040.png", "metadata/rollout_080/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_080/agentview/0041.png", "metadata/rollout_080/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_080/agentview/0041.png", "metadata/rollout_080/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0000.png", "metadata/rollout_081/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0001.png", "metadata/rollout_081/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0002.png", "metadata/rollout_081/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0003.png", "metadata/rollout_081/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0004.png", "metadata/rollout_081/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0005.png", "metadata/rollout_081/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0006.png", "metadata/rollout_081/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0007.png", "metadata/rollout_081/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0008.png", "metadata/rollout_081/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_081/agentview/0009.png", "metadata/rollout_081/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0010.png", "metadata/rollout_081/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0011.png", "metadata/rollout_081/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0012.png", "metadata/rollout_081/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0013.png", "metadata/rollout_081/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0014.png", "metadata/rollout_081/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0015.png", "metadata/rollout_081/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0016.png", "metadata/rollout_081/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0017.png", "metadata/rollout_081/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0018.png", "metadata/rollout_081/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0019.png", "metadata/rollout_081/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0020.png", "metadata/rollout_081/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0021.png", "metadata/rollout_081/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0022.png", "metadata/rollout_081/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0023.png", "metadata/rollout_081/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_081/agentview/0024.png", "metadata/rollout_081/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0025.png", "metadata/rollout_081/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0026.png", "metadata/rollout_081/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0027.png", "metadata/rollout_081/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0028.png", "metadata/rollout_081/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0029.png", "metadata/rollout_081/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0030.png", "metadata/rollout_081/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0031.png", "metadata/rollout_081/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0032.png", "metadata/rollout_081/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_081/agentview/0033.png", "metadata/rollout_081/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_081/agentview/0034.png", "metadata/rollout_081/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_081/agentview/0035.png", "metadata/rollout_081/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_081/agentview/0036.png", "metadata/rollout_081/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_081/agentview/0037.png", "metadata/rollout_081/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_081/agentview/0038.png", "metadata/rollout_081/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_081/agentview/0039.png", "metadata/rollout_081/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_081/agentview/0040.png", "metadata/rollout_081/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_081/agentview/0041.png", "metadata/rollout_081/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0042.png", "metadata/rollout_081/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0043.png", "metadata/rollout_081/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_081/agentview/0044.png", "metadata/rollout_081/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_081/agentview/0045.png", "metadata/rollout_081/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0046.png", "metadata/rollout_081/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_081/agentview/0047.png", "metadata/rollout_081/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_081/agentview/0047.png", "metadata/rollout_081/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0000.png", "metadata/rollout_082/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0001.png", "metadata/rollout_082/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0002.png", "metadata/rollout_082/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0003.png", "metadata/rollout_082/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0004.png", "metadata/rollout_082/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0005.png", "metadata/rollout_082/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0006.png", "metadata/rollout_082/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0007.png", "metadata/rollout_082/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0008.png", "metadata/rollout_082/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_082/agentview/0009.png", "metadata/rollout_082/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_082/agentview/0010.png", "metadata/rollout_082/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_082/agentview/0011.png", "metadata/rollout_082/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_082/agentview/0012.png", "metadata/rollout_082/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_082/agentview/0013.png", "metadata/rollout_082/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_082/agentview/0014.png", "metadata/rollout_082/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_082/agentview/0015.png", "metadata/rollout_082/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0016.png", "metadata/rollout_082/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0017.png", "metadata/rollout_082/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0018.png", "metadata/rollout_082/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0019.png", "metadata/rollout_082/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0020.png", "metadata/rollout_082/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_082/agentview/0021.png", "metadata/rollout_082/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_082/agentview/0022.png", "metadata/rollout_082/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_082/agentview/0023.png", "metadata/rollout_082/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_082/agentview/0024.png", "metadata/rollout_082/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_082/agentview/0025.png", "metadata/rollout_082/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_082/agentview/0026.png", "metadata/rollout_082/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_082/agentview/0027.png", "metadata/rollout_082/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_082/agentview/0028.png", "metadata/rollout_082/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_082/agentview/0029.png", "metadata/rollout_082/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_082/agentview/0030.png", "metadata/rollout_082/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_082/agentview/0031.png", "metadata/rollout_082/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_082/agentview/0032.png", "metadata/rollout_082/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_082/agentview/0033.png", "metadata/rollout_082/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_082/agentview/0034.png", "metadata/rollout_082/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_082/agentview/0035.png", "metadata/rollout_082/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_082/agentview/0036.png", "metadata/rollout_082/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_082/agentview/0037.png", "metadata/rollout_082/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_082/agentview/0038.png", "metadata/rollout_082/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_082/agentview/0039.png", "metadata/rollout_082/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_082/agentview/0040.png", "metadata/rollout_082/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_082/agentview/0041.png", "metadata/rollout_082/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0042.png", "metadata/rollout_082/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_082/agentview/0043.png", "metadata/rollout_082/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_082/agentview/0044.png", "metadata/rollout_082/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_082/agentview/0045.png", "metadata/rollout_082/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_082/agentview/0046.png", "metadata/rollout_082/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_082/agentview/0046.png", "metadata/rollout_082/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0000.png", "metadata/rollout_083/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0001.png", "metadata/rollout_083/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0002.png", "metadata/rollout_083/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0003.png", "metadata/rollout_083/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0004.png", "metadata/rollout_083/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0005.png", "metadata/rollout_083/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0006.png", "metadata/rollout_083/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0007.png", "metadata/rollout_083/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0008.png", "metadata/rollout_083/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_083/agentview/0009.png", "metadata/rollout_083/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_083/agentview/0010.png", "metadata/rollout_083/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_083/agentview/0011.png", "metadata/rollout_083/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_083/agentview/0012.png", "metadata/rollout_083/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_083/agentview/0013.png", "metadata/rollout_083/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_083/agentview/0014.png", "metadata/rollout_083/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_083/agentview/0015.png", "metadata/rollout_083/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_083/agentview/0016.png", "metadata/rollout_083/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_083/agentview/0017.png", "metadata/rollout_083/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_083/agentview/0018.png", "metadata/rollout_083/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_083/agentview/0019.png", "metadata/rollout_083/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_083/agentview/0020.png", "metadata/rollout_083/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_083/agentview/0021.png", "metadata/rollout_083/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_083/agentview/0022.png", "metadata/rollout_083/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_083/agentview/0023.png", "metadata/rollout_083/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0024.png", "metadata/rollout_083/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0025.png", "metadata/rollout_083/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0026.png", "metadata/rollout_083/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0027.png", "metadata/rollout_083/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0028.png", "metadata/rollout_083/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_083/agentview/0029.png", "metadata/rollout_083/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0030.png", "metadata/rollout_083/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0031.png", "metadata/rollout_083/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0032.png", "metadata/rollout_083/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0033.png", "metadata/rollout_083/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0034.png", "metadata/rollout_083/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0035.png", "metadata/rollout_083/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0036.png", "metadata/rollout_083/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0037.png", "metadata/rollout_083/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_083/agentview/0038.png", "metadata/rollout_083/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_083/agentview/0039.png", "metadata/rollout_083/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0040.png", "metadata/rollout_083/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0041.png", "metadata/rollout_083/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0042.png", "metadata/rollout_083/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0043.png", "metadata/rollout_083/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0044.png", "metadata/rollout_083/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0045.png", "metadata/rollout_083/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0046.png", "metadata/rollout_083/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0047.png", "metadata/rollout_083/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0048.png", "metadata/rollout_083/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0049.png", "metadata/rollout_083/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_083/agentview/0050.png", "metadata/rollout_083/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0051.png", "metadata/rollout_083/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_083/agentview/0052.png", "metadata/rollout_083/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_083/agentview/0053.png", "metadata/rollout_083/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0054.png", "metadata/rollout_083/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_083/agentview/0055.png", "metadata/rollout_083/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_083/agentview/0055.png", "metadata/rollout_083/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0000.png", "metadata/rollout_084/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0001.png", "metadata/rollout_084/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0002.png", "metadata/rollout_084/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0003.png", "metadata/rollout_084/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0004.png", "metadata/rollout_084/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0005.png", "metadata/rollout_084/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0006.png", "metadata/rollout_084/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0007.png", "metadata/rollout_084/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0008.png", "metadata/rollout_084/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0009.png", "metadata/rollout_084/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0010.png", "metadata/rollout_084/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0011.png", "metadata/rollout_084/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0012.png", "metadata/rollout_084/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0013.png", "metadata/rollout_084/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0014.png", "metadata/rollout_084/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0015.png", "metadata/rollout_084/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0016.png", "metadata/rollout_084/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0017.png", "metadata/rollout_084/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_084/agentview/0018.png", "metadata/rollout_084/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_084/agentview/0019.png", "metadata/rollout_084/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_084/agentview/0020.png", "metadata/rollout_084/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_084/agentview/0021.png", "metadata/rollout_084/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_084/agentview/0022.png", "metadata/rollout_084/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0023.png", "metadata/rollout_084/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0024.png", "metadata/rollout_084/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0025.png", "metadata/rollout_084/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0026.png", "metadata/rollout_084/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0027.png", "metadata/rollout_084/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_084/agentview/0028.png", "metadata/rollout_084/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0029.png", "metadata/rollout_084/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0030.png", "metadata/rollout_084/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0031.png", "metadata/rollout_084/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0032.png", "metadata/rollout_084/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0033.png", "metadata/rollout_084/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0034.png", "metadata/rollout_084/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0035.png", "metadata/rollout_084/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0036.png", "metadata/rollout_084/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0037.png", "metadata/rollout_084/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0038.png", "metadata/rollout_084/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0039.png", "metadata/rollout_084/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0040.png", "metadata/rollout_084/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0041.png", "metadata/rollout_084/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0042.png", "metadata/rollout_084/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0043.png", "metadata/rollout_084/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0044.png", "metadata/rollout_084/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0045.png", "metadata/rollout_084/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0046.png", "metadata/rollout_084/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0047.png", "metadata/rollout_084/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0048.png", "metadata/rollout_084/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_084/agentview/0049.png", "metadata/rollout_084/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_084/agentview/0050.png", "metadata/rollout_084/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0051.png", "metadata/rollout_084/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_084/agentview/0052.png", "metadata/rollout_084/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_084/agentview/0053.png", "metadata/rollout_084/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0054.png", "metadata/rollout_084/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_084/agentview/0055.png", "metadata/rollout_084/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_084/agentview/0055.png", "metadata/rollout_084/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0000.png", "metadata/rollout_085/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0001.png", "metadata/rollout_085/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0002.png", "metadata/rollout_085/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0003.png", "metadata/rollout_085/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0004.png", "metadata/rollout_085/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0005.png", "metadata/rollout_085/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0006.png", "metadata/rollout_085/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0007.png", "metadata/rollout_085/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0008.png", "metadata/rollout_085/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0009.png", "metadata/rollout_085/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0010.png", "metadata/rollout_085/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0011.png", "metadata/rollout_085/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0012.png", "metadata/rollout_085/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0013.png", "metadata/rollout_085/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0014.png", "metadata/rollout_085/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0015.png", "metadata/rollout_085/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0016.png", "metadata/rollout_085/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0017.png", "metadata/rollout_085/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_085/agentview/0018.png", "metadata/rollout_085/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0019.png", "metadata/rollout_085/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0020.png", "metadata/rollout_085/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0021.png", "metadata/rollout_085/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0022.png", "metadata/rollout_085/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0023.png", "metadata/rollout_085/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0024.png", "metadata/rollout_085/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_085/agentview/0025.png", "metadata/rollout_085/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0026.png", "metadata/rollout_085/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0027.png", "metadata/rollout_085/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0028.png", "metadata/rollout_085/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0029.png", "metadata/rollout_085/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0030.png", "metadata/rollout_085/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0031.png", "metadata/rollout_085/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0032.png", "metadata/rollout_085/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0033.png", "metadata/rollout_085/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_085/agentview/0034.png", "metadata/rollout_085/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_085/agentview/0035.png", "metadata/rollout_085/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_085/agentview/0036.png", "metadata/rollout_085/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_085/agentview/0037.png", "metadata/rollout_085/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_085/agentview/0038.png", "metadata/rollout_085/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_085/agentview/0039.png", "metadata/rollout_085/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_085/agentview/0040.png", "metadata/rollout_085/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_085/agentview/0041.png", "metadata/rollout_085/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0042.png", "metadata/rollout_085/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0043.png", "metadata/rollout_085/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_085/agentview/0044.png", "metadata/rollout_085/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_085/agentview/0045.png", "metadata/rollout_085/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0046.png", "metadata/rollout_085/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_085/agentview/0047.png", "metadata/rollout_085/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_085/agentview/0047.png", "metadata/rollout_085/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0000.png", "metadata/rollout_086/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0001.png", "metadata/rollout_086/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0002.png", "metadata/rollout_086/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0003.png", "metadata/rollout_086/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0004.png", "metadata/rollout_086/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0005.png", "metadata/rollout_086/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0006.png", "metadata/rollout_086/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0007.png", "metadata/rollout_086/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0008.png", "metadata/rollout_086/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_086/agentview/0009.png", "metadata/rollout_086/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_086/agentview/0010.png", "metadata/rollout_086/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_086/agentview/0011.png", "metadata/rollout_086/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0012.png", "metadata/rollout_086/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0013.png", "metadata/rollout_086/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0014.png", "metadata/rollout_086/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0015.png", "metadata/rollout_086/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0016.png", "metadata/rollout_086/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_086/agentview/0017.png", "metadata/rollout_086/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_086/agentview/0018.png", "metadata/rollout_086/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_086/agentview/0019.png", "metadata/rollout_086/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_086/agentview/0020.png", "metadata/rollout_086/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_086/agentview/0021.png", "metadata/rollout_086/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_086/agentview/0022.png", "metadata/rollout_086/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_086/agentview/0023.png", "metadata/rollout_086/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_086/agentview/0024.png", "metadata/rollout_086/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_086/agentview/0025.png", "metadata/rollout_086/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_086/agentview/0026.png", "metadata/rollout_086/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_086/agentview/0027.png", "metadata/rollout_086/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_086/agentview/0028.png", "metadata/rollout_086/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_086/agentview/0029.png", "metadata/rollout_086/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_086/agentview/0030.png", "metadata/rollout_086/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_086/agentview/0031.png", "metadata/rollout_086/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_086/agentview/0032.png", "metadata/rollout_086/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_086/agentview/0033.png", "metadata/rollout_086/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_086/agentview/0034.png", "metadata/rollout_086/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_086/agentview/0035.png", "metadata/rollout_086/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0036.png", "metadata/rollout_086/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_086/agentview/0037.png", "metadata/rollout_086/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_086/agentview/0038.png", "metadata/rollout_086/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_086/agentview/0039.png", "metadata/rollout_086/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_086/agentview/0040.png", "metadata/rollout_086/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_086/agentview/0040.png", "metadata/rollout_086/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0000.png", "metadata/rollout_087/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0001.png", "metadata/rollout_087/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0002.png", "metadata/rollout_087/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0003.png", "metadata/rollout_087/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0004.png", "metadata/rollout_087/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0005.png", "metadata/rollout_087/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0006.png", "metadata/rollout_087/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0007.png", "metadata/rollout_087/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0008.png", "metadata/rollout_087/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0009.png", "metadata/rollout_087/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0010.png", "metadata/rollout_087/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0011.png", "metadata/rollout_087/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0012.png", "metadata/rollout_087/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0013.png", "metadata/rollout_087/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0014.png", "metadata/rollout_087/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0015.png", "metadata/rollout_087/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0016.png", "metadata/rollout_087/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0017.png", "metadata/rollout_087/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0018.png", "metadata/rollout_087/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_087/agentview/0019.png", "metadata/rollout_087/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_087/agentview/0020.png", "metadata/rollout_087/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_087/agentview/0021.png", "metadata/rollout_087/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_087/agentview/0022.png", "metadata/rollout_087/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0023.png", "metadata/rollout_087/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0024.png", "metadata/rollout_087/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0025.png", "metadata/rollout_087/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0026.png", "metadata/rollout_087/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0027.png", "metadata/rollout_087/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_087/agentview/0028.png", "metadata/rollout_087/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0029.png", "metadata/rollout_087/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0030.png", "metadata/rollout_087/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0031.png", "metadata/rollout_087/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0032.png", "metadata/rollout_087/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0033.png", "metadata/rollout_087/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0034.png", "metadata/rollout_087/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0035.png", "metadata/rollout_087/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0036.png", "metadata/rollout_087/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0037.png", "metadata/rollout_087/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0038.png", "metadata/rollout_087/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0039.png", "metadata/rollout_087/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0040.png", "metadata/rollout_087/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0041.png", "metadata/rollout_087/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0042.png", "metadata/rollout_087/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0043.png", "metadata/rollout_087/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0044.png", "metadata/rollout_087/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0045.png", "metadata/rollout_087/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0046.png", "metadata/rollout_087/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0047.png", "metadata/rollout_087/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0048.png", "metadata/rollout_087/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_087/agentview/0049.png", "metadata/rollout_087/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_087/agentview/0050.png", "metadata/rollout_087/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_087/agentview/0051.png", "metadata/rollout_087/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_087/agentview/0052.png", "metadata/rollout_087/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_087/agentview/0053.png", "metadata/rollout_087/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_087/agentview/0054.png", "metadata/rollout_087/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0055.png", "metadata/rollout_087/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_087/agentview/0056.png", "metadata/rollout_087/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_087/agentview/0057.png", "metadata/rollout_087/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0058.png", "metadata/rollout_087/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_087/agentview/0059.png", "metadata/rollout_087/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_087/agentview/0059.png", "metadata/rollout_087/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0000.png", "metadata/rollout_088/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0001.png", "metadata/rollout_088/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0002.png", "metadata/rollout_088/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0003.png", "metadata/rollout_088/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0004.png", "metadata/rollout_088/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0005.png", "metadata/rollout_088/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0006.png", "metadata/rollout_088/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0007.png", "metadata/rollout_088/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0008.png", "metadata/rollout_088/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0009.png", "metadata/rollout_088/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0010.png", "metadata/rollout_088/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0011.png", "metadata/rollout_088/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0012.png", "metadata/rollout_088/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0013.png", "metadata/rollout_088/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0014.png", "metadata/rollout_088/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0015.png", "metadata/rollout_088/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0016.png", "metadata/rollout_088/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0017.png", "metadata/rollout_088/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_088/agentview/0018.png", "metadata/rollout_088/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_088/agentview/0019.png", "metadata/rollout_088/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_088/agentview/0020.png", "metadata/rollout_088/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_088/agentview/0021.png", "metadata/rollout_088/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_088/agentview/0022.png", "metadata/rollout_088/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_088/agentview/0023.png", "metadata/rollout_088/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_088/agentview/0024.png", "metadata/rollout_088/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0025.png", "metadata/rollout_088/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0026.png", "metadata/rollout_088/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0027.png", "metadata/rollout_088/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0028.png", "metadata/rollout_088/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0029.png", "metadata/rollout_088/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_088/agentview/0030.png", "metadata/rollout_088/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0031.png", "metadata/rollout_088/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0032.png", "metadata/rollout_088/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0033.png", "metadata/rollout_088/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0034.png", "metadata/rollout_088/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0035.png", "metadata/rollout_088/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0036.png", "metadata/rollout_088/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0037.png", "metadata/rollout_088/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0038.png", "metadata/rollout_088/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0039.png", "metadata/rollout_088/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0040.png", "metadata/rollout_088/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0041.png", "metadata/rollout_088/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0042.png", "metadata/rollout_088/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0043.png", "metadata/rollout_088/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0044.png", "metadata/rollout_088/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0045.png", "metadata/rollout_088/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0046.png", "metadata/rollout_088/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0047.png", "metadata/rollout_088/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0048.png", "metadata/rollout_088/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0049.png", "metadata/rollout_088/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0050.png", "metadata/rollout_088/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_088/agentview/0051.png", "metadata/rollout_088/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_088/agentview/0052.png", "metadata/rollout_088/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0053.png", "metadata/rollout_088/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_088/agentview/0054.png", "metadata/rollout_088/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_088/agentview/0055.png", "metadata/rollout_088/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0056.png", "metadata/rollout_088/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_088/agentview/0057.png", "metadata/rollout_088/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_088/agentview/0057.png", "metadata/rollout_088/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0000.png", "metadata/rollout_089/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0001.png", "metadata/rollout_089/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0002.png", "metadata/rollout_089/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0003.png", "metadata/rollout_089/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0004.png", "metadata/rollout_089/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0005.png", "metadata/rollout_089/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0006.png", "metadata/rollout_089/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0007.png", "metadata/rollout_089/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0008.png", "metadata/rollout_089/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0009.png", "metadata/rollout_089/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0010.png", "metadata/rollout_089/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0011.png", "metadata/rollout_089/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0012.png", "metadata/rollout_089/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0013.png", "metadata/rollout_089/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0014.png", "metadata/rollout_089/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0015.png", "metadata/rollout_089/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0016.png", "metadata/rollout_089/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0017.png", "metadata/rollout_089/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_089/agentview/0018.png", "metadata/rollout_089/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0019.png", "metadata/rollout_089/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0020.png", "metadata/rollout_089/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0021.png", "metadata/rollout_089/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0022.png", "metadata/rollout_089/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0023.png", "metadata/rollout_089/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_089/agentview/0024.png", "metadata/rollout_089/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0025.png", "metadata/rollout_089/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0026.png", "metadata/rollout_089/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0027.png", "metadata/rollout_089/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0028.png", "metadata/rollout_089/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0029.png", "metadata/rollout_089/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0030.png", "metadata/rollout_089/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0031.png", "metadata/rollout_089/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0032.png", "metadata/rollout_089/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_089/agentview/0033.png", "metadata/rollout_089/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_089/agentview/0034.png", "metadata/rollout_089/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_089/agentview/0035.png", "metadata/rollout_089/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_089/agentview/0036.png", "metadata/rollout_089/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_089/agentview/0037.png", "metadata/rollout_089/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_089/agentview/0038.png", "metadata/rollout_089/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_089/agentview/0039.png", "metadata/rollout_089/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_089/agentview/0040.png", "metadata/rollout_089/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_089/agentview/0041.png", "metadata/rollout_089/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0042.png", "metadata/rollout_089/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0043.png", "metadata/rollout_089/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_089/agentview/0044.png", "metadata/rollout_089/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0045.png", "metadata/rollout_089/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_089/agentview/0046.png", "metadata/rollout_089/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0047.png", "metadata/rollout_089/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_089/agentview/0048.png", "metadata/rollout_089/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the cube in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_089/agentview/0048.png", "metadata/rollout_089/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0000.png", "metadata/rollout_090/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0001.png", "metadata/rollout_090/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0002.png", "metadata/rollout_090/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0003.png", "metadata/rollout_090/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0004.png", "metadata/rollout_090/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0005.png", "metadata/rollout_090/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0006.png", "metadata/rollout_090/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0007.png", "metadata/rollout_090/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0008.png", "metadata/rollout_090/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0009.png", "metadata/rollout_090/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0010.png", "metadata/rollout_090/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0011.png", "metadata/rollout_090/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0012.png", "metadata/rollout_090/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_090/agentview/0013.png", "metadata/rollout_090/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_090/agentview/0014.png", "metadata/rollout_090/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0015.png", "metadata/rollout_090/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0016.png", "metadata/rollout_090/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0017.png", "metadata/rollout_090/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0018.png", "metadata/rollout_090/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0019.png", "metadata/rollout_090/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0020.png", "metadata/rollout_090/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_090/agentview/0021.png", "metadata/rollout_090/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0022.png", "metadata/rollout_090/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0023.png", "metadata/rollout_090/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0024.png", "metadata/rollout_090/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0025.png", "metadata/rollout_090/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0026.png", "metadata/rollout_090/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0027.png", "metadata/rollout_090/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0028.png", "metadata/rollout_090/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0029.png", "metadata/rollout_090/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_090/agentview/0030.png", "metadata/rollout_090/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_090/agentview/0031.png", "metadata/rollout_090/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_090/agentview/0032.png", "metadata/rollout_090/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_090/agentview/0033.png", "metadata/rollout_090/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_090/agentview/0034.png", "metadata/rollout_090/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0035.png", "metadata/rollout_090/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0036.png", "metadata/rollout_090/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0037.png", "metadata/rollout_090/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0038.png", "metadata/rollout_090/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0039.png", "metadata/rollout_090/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_090/agentview/0040.png", "metadata/rollout_090/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0041.png", "metadata/rollout_090/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_090/agentview/0042.png", "metadata/rollout_090/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0043.png", "metadata/rollout_090/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_090/agentview/0044.png", "metadata/rollout_090/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0045.png", "metadata/rollout_090/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_090/agentview/0046.png", "metadata/rollout_090/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_090/agentview/0046.png", "metadata/rollout_090/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0000.png", "metadata/rollout_091/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0001.png", "metadata/rollout_091/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0002.png", "metadata/rollout_091/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0003.png", "metadata/rollout_091/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0004.png", "metadata/rollout_091/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0005.png", "metadata/rollout_091/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0006.png", "metadata/rollout_091/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0007.png", "metadata/rollout_091/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0008.png", "metadata/rollout_091/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0009.png", "metadata/rollout_091/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0010.png", "metadata/rollout_091/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0011.png", "metadata/rollout_091/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0012.png", "metadata/rollout_091/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0013.png", "metadata/rollout_091/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0014.png", "metadata/rollout_091/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0015.png", "metadata/rollout_091/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0016.png", "metadata/rollout_091/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0017.png", "metadata/rollout_091/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0018.png", "metadata/rollout_091/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_091/agentview/0019.png", "metadata/rollout_091/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0020.png", "metadata/rollout_091/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0021.png", "metadata/rollout_091/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0022.png", "metadata/rollout_091/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0023.png", "metadata/rollout_091/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0024.png", "metadata/rollout_091/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0025.png", "metadata/rollout_091/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0026.png", "metadata/rollout_091/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_091/agentview/0027.png", "metadata/rollout_091/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_091/agentview/0028.png", "metadata/rollout_091/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_091/agentview/0029.png", "metadata/rollout_091/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0030.png", "metadata/rollout_091/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0031.png", "metadata/rollout_091/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0032.png", "metadata/rollout_091/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0033.png", "metadata/rollout_091/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0034.png", "metadata/rollout_091/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0035.png", "metadata/rollout_091/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0036.png", "metadata/rollout_091/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_091/agentview/0037.png", "metadata/rollout_091/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_091/agentview/0038.png", "metadata/rollout_091/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_091/agentview/0039.png", "metadata/rollout_091/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_091/agentview/0040.png", "metadata/rollout_091/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_091/agentview/0041.png", "metadata/rollout_091/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_091/agentview/0042.png", "metadata/rollout_091/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_091/agentview/0043.png", "metadata/rollout_091/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_091/agentview/0044.png", "metadata/rollout_091/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_091/agentview/0045.png", "metadata/rollout_091/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_091/agentview/0046.png", "metadata/rollout_091/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_091/agentview/0047.png", "metadata/rollout_091/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_091/agentview/0048.png", "metadata/rollout_091/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_091/agentview/0049.png", "metadata/rollout_091/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_091/agentview/0050.png", "metadata/rollout_091/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_091/agentview/0051.png", "metadata/rollout_091/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0052.png", "metadata/rollout_091/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_091/agentview/0053.png", "metadata/rollout_091/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0054.png", "metadata/rollout_091/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_091/agentview/0055.png", "metadata/rollout_091/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_091/agentview/0055.png", "metadata/rollout_091/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0000.png", "metadata/rollout_092/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0001.png", "metadata/rollout_092/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0002.png", "metadata/rollout_092/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0003.png", "metadata/rollout_092/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0004.png", "metadata/rollout_092/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0005.png", "metadata/rollout_092/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0006.png", "metadata/rollout_092/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0007.png", "metadata/rollout_092/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_092/agentview/0008.png", "metadata/rollout_092/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_092/agentview/0009.png", "metadata/rollout_092/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_092/agentview/0010.png", "metadata/rollout_092/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_092/agentview/0011.png", "metadata/rollout_092/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_092/agentview/0012.png", "metadata/rollout_092/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_092/agentview/0013.png", "metadata/rollout_092/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_092/agentview/0014.png", "metadata/rollout_092/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_092/agentview/0015.png", "metadata/rollout_092/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_092/agentview/0016.png", "metadata/rollout_092/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0017.png", "metadata/rollout_092/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0018.png", "metadata/rollout_092/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0019.png", "metadata/rollout_092/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0020.png", "metadata/rollout_092/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0021.png", "metadata/rollout_092/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0022.png", "metadata/rollout_092/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_092/agentview/0023.png", "metadata/rollout_092/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0024.png", "metadata/rollout_092/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0025.png", "metadata/rollout_092/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0026.png", "metadata/rollout_092/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0027.png", "metadata/rollout_092/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0028.png", "metadata/rollout_092/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0029.png", "metadata/rollout_092/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0030.png", "metadata/rollout_092/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0031.png", "metadata/rollout_092/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_092/agentview/0032.png", "metadata/rollout_092/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_092/agentview/0033.png", "metadata/rollout_092/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_092/agentview/0034.png", "metadata/rollout_092/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0035.png", "metadata/rollout_092/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0036.png", "metadata/rollout_092/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0037.png", "metadata/rollout_092/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0038.png", "metadata/rollout_092/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0039.png", "metadata/rollout_092/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0040.png", "metadata/rollout_092/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0041.png", "metadata/rollout_092/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0042.png", "metadata/rollout_092/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0043.png", "metadata/rollout_092/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_092/agentview/0044.png", "metadata/rollout_092/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_092/agentview/0045.png", "metadata/rollout_092/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_092/agentview/0046.png", "metadata/rollout_092/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_092/agentview/0047.png", "metadata/rollout_092/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0048.png", "metadata/rollout_092/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_092/agentview/0049.png", "metadata/rollout_092/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_092/agentview/0049.png", "metadata/rollout_092/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_093/agentview/0000.png", "metadata/rollout_093/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_093/agentview/0001.png", "metadata/rollout_093/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_093/agentview/0002.png", "metadata/rollout_093/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_093/agentview/0003.png", "metadata/rollout_093/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_093/agentview/0004.png", "metadata/rollout_093/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_093/agentview/0005.png", "metadata/rollout_093/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_093/agentview/0006.png", "metadata/rollout_093/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_093/agentview/0007.png", "metadata/rollout_093/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_093/agentview/0008.png", "metadata/rollout_093/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0009.png", "metadata/rollout_093/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0010.png", "metadata/rollout_093/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0011.png", "metadata/rollout_093/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0012.png", "metadata/rollout_093/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0013.png", "metadata/rollout_093/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0014.png", "metadata/rollout_093/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0015.png", "metadata/rollout_093/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0016.png", "metadata/rollout_093/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0017.png", "metadata/rollout_093/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0018.png", "metadata/rollout_093/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0019.png", "metadata/rollout_093/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0020.png", "metadata/rollout_093/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0021.png", "metadata/rollout_093/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0022.png", "metadata/rollout_093/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0023.png", "metadata/rollout_093/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0024.png", "metadata/rollout_093/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0025.png", "metadata/rollout_093/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0026.png", "metadata/rollout_093/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0027.png", "metadata/rollout_093/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0028.png", "metadata/rollout_093/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0029.png", "metadata/rollout_093/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0030.png", "metadata/rollout_093/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0031.png", "metadata/rollout_093/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_093/agentview/0032.png", "metadata/rollout_093/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0033.png", "metadata/rollout_093/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0034.png", "metadata/rollout_093/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0035.png", "metadata/rollout_093/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0036.png", "metadata/rollout_093/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0037.png", "metadata/rollout_093/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0038.png", "metadata/rollout_093/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0039.png", "metadata/rollout_093/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0040.png", "metadata/rollout_093/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_093/agentview/0041.png", "metadata/rollout_093/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_093/agentview/0042.png", "metadata/rollout_093/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_093/agentview/0043.png", "metadata/rollout_093/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0044.png", "metadata/rollout_093/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0045.png", "metadata/rollout_093/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0046.png", "metadata/rollout_093/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0047.png", "metadata/rollout_093/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0048.png", "metadata/rollout_093/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0049.png", "metadata/rollout_093/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0050.png", "metadata/rollout_093/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0051.png", "metadata/rollout_093/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0052.png", "metadata/rollout_093/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0053.png", "metadata/rollout_093/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0054.png", "metadata/rollout_093/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_093/agentview/0055.png", "metadata/rollout_093/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0056.png", "metadata/rollout_093/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_093/agentview/0057.png", "metadata/rollout_093/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_093/agentview/0058.png", "metadata/rollout_093/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_093/agentview/0059.png", "metadata/rollout_093/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0060.png", "metadata/rollout_093/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_093/agentview/0061.png", "metadata/rollout_093/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_093/agentview/0061.png", "metadata/rollout_093/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0000.png", "metadata/rollout_094/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0001.png", "metadata/rollout_094/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0002.png", "metadata/rollout_094/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0003.png", "metadata/rollout_094/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0004.png", "metadata/rollout_094/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0005.png", "metadata/rollout_094/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0006.png", "metadata/rollout_094/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0007.png", "metadata/rollout_094/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0008.png", "metadata/rollout_094/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0009.png", "metadata/rollout_094/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0010.png", "metadata/rollout_094/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0011.png", "metadata/rollout_094/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0012.png", "metadata/rollout_094/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0013.png", "metadata/rollout_094/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0014.png", "metadata/rollout_094/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0015.png", "metadata/rollout_094/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0016.png", "metadata/rollout_094/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_094/agentview/0017.png", "metadata/rollout_094/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_094/agentview/0018.png", "metadata/rollout_094/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_094/agentview/0019.png", "metadata/rollout_094/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_094/agentview/0020.png", "metadata/rollout_094/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_094/agentview/0021.png", "metadata/rollout_094/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_094/agentview/0022.png", "metadata/rollout_094/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_094/agentview/0023.png", "metadata/rollout_094/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0024.png", "metadata/rollout_094/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0025.png", "metadata/rollout_094/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0026.png", "metadata/rollout_094/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0027.png", "metadata/rollout_094/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0028.png", "metadata/rollout_094/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_094/agentview/0029.png", "metadata/rollout_094/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0030.png", "metadata/rollout_094/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_094/agentview/0031.png", "metadata/rollout_094/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0032.png", "metadata/rollout_094/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0033.png", "metadata/rollout_094/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0034.png", "metadata/rollout_094/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0035.png", "metadata/rollout_094/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0036.png", "metadata/rollout_094/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0037.png", "metadata/rollout_094/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0038.png", "metadata/rollout_094/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0039.png", "metadata/rollout_094/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0040.png", "metadata/rollout_094/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0041.png", "metadata/rollout_094/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0042.png", "metadata/rollout_094/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0043.png", "metadata/rollout_094/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0044.png", "metadata/rollout_094/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0045.png", "metadata/rollout_094/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0046.png", "metadata/rollout_094/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0047.png", "metadata/rollout_094/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0048.png", "metadata/rollout_094/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0049.png", "metadata/rollout_094/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0050.png", "metadata/rollout_094/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0051.png", "metadata/rollout_094/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0052.png", "metadata/rollout_094/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_094/agentview/0053.png", "metadata/rollout_094/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_094/agentview/0054.png", "metadata/rollout_094/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_094/agentview/0055.png", "metadata/rollout_094/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0056.png", "metadata/rollout_094/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_094/agentview/0057.png", "metadata/rollout_094/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_094/agentview/0057.png", "metadata/rollout_094/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_095/agentview/0000.png", "metadata/rollout_095/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_095/agentview/0001.png", "metadata/rollout_095/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_095/agentview/0002.png", "metadata/rollout_095/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_095/agentview/0003.png", "metadata/rollout_095/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_095/agentview/0004.png", "metadata/rollout_095/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_095/agentview/0005.png", "metadata/rollout_095/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_095/agentview/0006.png", "metadata/rollout_095/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_095/agentview/0007.png", "metadata/rollout_095/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_095/agentview/0008.png", "metadata/rollout_095/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0009.png", "metadata/rollout_095/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0010.png", "metadata/rollout_095/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0011.png", "metadata/rollout_095/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0012.png", "metadata/rollout_095/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0013.png", "metadata/rollout_095/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0014.png", "metadata/rollout_095/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0015.png", "metadata/rollout_095/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0016.png", "metadata/rollout_095/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_095/agentview/0017.png", "metadata/rollout_095/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_095/agentview/0018.png", "metadata/rollout_095/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_095/agentview/0019.png", "metadata/rollout_095/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0020.png", "metadata/rollout_095/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0021.png", "metadata/rollout_095/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0022.png", "metadata/rollout_095/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0023.png", "metadata/rollout_095/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0024.png", "metadata/rollout_095/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0025.png", "metadata/rollout_095/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_095/agentview/0026.png", "metadata/rollout_095/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0027.png", "metadata/rollout_095/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0028.png", "metadata/rollout_095/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0029.png", "metadata/rollout_095/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0030.png", "metadata/rollout_095/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0031.png", "metadata/rollout_095/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0032.png", "metadata/rollout_095/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0033.png", "metadata/rollout_095/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0034.png", "metadata/rollout_095/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_095/agentview/0035.png", "metadata/rollout_095/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_095/agentview/0036.png", "metadata/rollout_095/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_095/agentview/0037.png", "metadata/rollout_095/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_095/agentview/0038.png", "metadata/rollout_095/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_095/agentview/0039.png", "metadata/rollout_095/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_095/agentview/0040.png", "metadata/rollout_095/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_095/agentview/0041.png", "metadata/rollout_095/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_095/agentview/0042.png", "metadata/rollout_095/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0043.png", "metadata/rollout_095/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_095/agentview/0044.png", "metadata/rollout_095/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0045.png", "metadata/rollout_095/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_095/agentview/0046.png", "metadata/rollout_095/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0047.png", "metadata/rollout_095/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_095/agentview/0048.png", "metadata/rollout_095/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_095/agentview/0048.png", "metadata/rollout_095/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0000.png", "metadata/rollout_096/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0001.png", "metadata/rollout_096/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0002.png", "metadata/rollout_096/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0003.png", "metadata/rollout_096/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0004.png", "metadata/rollout_096/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0005.png", "metadata/rollout_096/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0006.png", "metadata/rollout_096/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0007.png", "metadata/rollout_096/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_096/agentview/0008.png", "metadata/rollout_096/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_096/agentview/0009.png", "metadata/rollout_096/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_096/agentview/0010.png", "metadata/rollout_096/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_096/agentview/0011.png", "metadata/rollout_096/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_096/agentview/0012.png", "metadata/rollout_096/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0013.png", "metadata/rollout_096/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0014.png", "metadata/rollout_096/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0015.png", "metadata/rollout_096/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0016.png", "metadata/rollout_096/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0017.png", "metadata/rollout_096/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0018.png", "metadata/rollout_096/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_096/agentview/0019.png", "metadata/rollout_096/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0020.png", "metadata/rollout_096/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0021.png", "metadata/rollout_096/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0022.png", "metadata/rollout_096/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0023.png", "metadata/rollout_096/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0024.png", "metadata/rollout_096/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0025.png", "metadata/rollout_096/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0026.png", "metadata/rollout_096/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0027.png", "metadata/rollout_096/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0028.png", "metadata/rollout_096/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0029.png", "metadata/rollout_096/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0030.png", "metadata/rollout_096/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0031.png", "metadata/rollout_096/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0032.png", "metadata/rollout_096/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0033.png", "metadata/rollout_096/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0034.png", "metadata/rollout_096/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0035.png", "metadata/rollout_096/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0036.png", "metadata/rollout_096/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_096/agentview/0037.png", "metadata/rollout_096/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_096/agentview/0038.png", "metadata/rollout_096/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_096/agentview/0039.png", "metadata/rollout_096/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0040.png", "metadata/rollout_096/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_096/agentview/0041.png", "metadata/rollout_096/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_096/agentview/0042.png", "metadata/rollout_096/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0043.png", "metadata/rollout_096/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_096/agentview/0044.png", "metadata/rollout_096/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_096/agentview/0044.png", "metadata/rollout_096/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0000.png", "metadata/rollout_097/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0001.png", "metadata/rollout_097/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0002.png", "metadata/rollout_097/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0003.png", "metadata/rollout_097/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0004.png", "metadata/rollout_097/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0005.png", "metadata/rollout_097/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0006.png", "metadata/rollout_097/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0007.png", "metadata/rollout_097/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0008.png", "metadata/rollout_097/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0009.png", "metadata/rollout_097/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0010.png", "metadata/rollout_097/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0011.png", "metadata/rollout_097/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0012.png", "metadata/rollout_097/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0013.png", "metadata/rollout_097/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0014.png", "metadata/rollout_097/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0015.png", "metadata/rollout_097/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0016.png", "metadata/rollout_097/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0017.png", "metadata/rollout_097/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0018.png", "metadata/rollout_097/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_097/agentview/0019.png", "metadata/rollout_097/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_097/agentview/0020.png", "metadata/rollout_097/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_097/agentview/0021.png", "metadata/rollout_097/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_097/agentview/0022.png", "metadata/rollout_097/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_097/agentview/0023.png", "metadata/rollout_097/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0024.png", "metadata/rollout_097/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0025.png", "metadata/rollout_097/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0026.png", "metadata/rollout_097/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0027.png", "metadata/rollout_097/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0028.png", "metadata/rollout_097/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0029.png", "metadata/rollout_097/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0030.png", "metadata/rollout_097/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0031.png", "metadata/rollout_097/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_097/agentview/0032.png", "metadata/rollout_097/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_097/agentview/0033.png", "metadata/rollout_097/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_097/agentview/0034.png", "metadata/rollout_097/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_097/agentview/0035.png", "metadata/rollout_097/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_097/agentview/0036.png", "metadata/rollout_097/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_097/agentview/0037.png", "metadata/rollout_097/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_097/agentview/0038.png", "metadata/rollout_097/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_097/agentview/0039.png", "metadata/rollout_097/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_097/agentview/0040.png", "metadata/rollout_097/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0041.png", "metadata/rollout_097/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0042.png", "metadata/rollout_097/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0043.png", "metadata/rollout_097/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0044.png", "metadata/rollout_097/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0045.png", "metadata/rollout_097/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0046.png", "metadata/rollout_097/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0047.png", "metadata/rollout_097/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0048.png", "metadata/rollout_097/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0049.png", "metadata/rollout_097/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0050.png", "metadata/rollout_097/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0051.png", "metadata/rollout_097/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0052.png", "metadata/rollout_097/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0053.png", "metadata/rollout_097/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0054.png", "metadata/rollout_097/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0055.png", "metadata/rollout_097/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0056.png", "metadata/rollout_097/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_097/agentview/0057.png", "metadata/rollout_097/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_097/agentview/0058.png", "metadata/rollout_097/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_097/agentview/0059.png", "metadata/rollout_097/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_097/agentview/0060.png", "metadata/rollout_097/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_097/agentview/0061.png", "metadata/rollout_097/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_097/agentview/0062.png", "metadata/rollout_097/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_097/agentview/0063.png", "metadata/rollout_097/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0064.png", "metadata/rollout_097/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_097/agentview/0065.png", "metadata/rollout_097/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_097/agentview/0066.png", "metadata/rollout_097/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_RIGHT, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_097/agentview/0067.png", "metadata/rollout_097/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_RIGHT, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_097/agentview/0068.png", "metadata/rollout_097/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_097/agentview/0069.png", "metadata/rollout_097/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_097/agentview/0069.png", "metadata/rollout_097/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0000.png", "metadata/rollout_098/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0001.png", "metadata/rollout_098/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0002.png", "metadata/rollout_098/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0003.png", "metadata/rollout_098/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0004.png", "metadata/rollout_098/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0005.png", "metadata/rollout_098/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0006.png", "metadata/rollout_098/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0007.png", "metadata/rollout_098/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0008.png", "metadata/rollout_098/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0009.png", "metadata/rollout_098/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_098/agentview/0010.png", "metadata/rollout_098/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_098/agentview/0011.png", "metadata/rollout_098/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_098/agentview/0012.png", "metadata/rollout_098/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_098/agentview/0013.png", "metadata/rollout_098/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_098/agentview/0014.png", "metadata/rollout_098/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_098/agentview/0015.png", "metadata/rollout_098/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_098/agentview/0016.png", "metadata/rollout_098/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_098/agentview/0017.png", "metadata/rollout_098/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_098/agentview/0018.png", "metadata/rollout_098/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0019.png", "metadata/rollout_098/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0020.png", "metadata/rollout_098/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0021.png", "metadata/rollout_098/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0022.png", "metadata/rollout_098/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0023.png", "metadata/rollout_098/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0024.png", "metadata/rollout_098/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0025.png", "metadata/rollout_098/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0026.png", "metadata/rollout_098/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0027.png", "metadata/rollout_098/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0028.png", "metadata/rollout_098/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0029.png", "metadata/rollout_098/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0030.png", "metadata/rollout_098/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0031.png", "metadata/rollout_098/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_098/agentview/0032.png", "metadata/rollout_098/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_098/agentview/0033.png", "metadata/rollout_098/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0034.png", "metadata/rollout_098/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0035.png", "metadata/rollout_098/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0036.png", "metadata/rollout_098/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0037.png", "metadata/rollout_098/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0038.png", "metadata/rollout_098/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0039.png", "metadata/rollout_098/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0040.png", "metadata/rollout_098/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0041.png", "metadata/rollout_098/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0042.png", "metadata/rollout_098/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0043.png", "metadata/rollout_098/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0044.png", "metadata/rollout_098/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0045.png", "metadata/rollout_098/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0046.png", "metadata/rollout_098/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0047.png", "metadata/rollout_098/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0048.png", "metadata/rollout_098/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0049.png", "metadata/rollout_098/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0050.png", "metadata/rollout_098/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0051.png", "metadata/rollout_098/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0052.png", "metadata/rollout_098/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0053.png", "metadata/rollout_098/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0054.png", "metadata/rollout_098/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0055.png", "metadata/rollout_098/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_098/agentview/0056.png", "metadata/rollout_098/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0057.png", "metadata/rollout_098/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_098/agentview/0058.png", "metadata/rollout_098/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_098/agentview/0059.png", "metadata/rollout_098/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0060.png", "metadata/rollout_098/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_098/agentview/0061.png", "metadata/rollout_098/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_098/agentview/0061.png", "metadata/rollout_098/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0000.png", "metadata/rollout_099/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0001.png", "metadata/rollout_099/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0002.png", "metadata/rollout_099/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0003.png", "metadata/rollout_099/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0004.png", "metadata/rollout_099/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0005.png", "metadata/rollout_099/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0006.png", "metadata/rollout_099/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0007.png", "metadata/rollout_099/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0008.png", "metadata/rollout_099/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0009.png", "metadata/rollout_099/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0010.png", "metadata/rollout_099/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0011.png", "metadata/rollout_099/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0012.png", "metadata/rollout_099/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0013.png", "metadata/rollout_099/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0014.png", "metadata/rollout_099/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0015.png", "metadata/rollout_099/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0016.png", "metadata/rollout_099/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0017.png", "metadata/rollout_099/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_099/agentview/0018.png", "metadata/rollout_099/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_099/agentview/0019.png", "metadata/rollout_099/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_099/agentview/0020.png", "metadata/rollout_099/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0021.png", "metadata/rollout_099/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0022.png", "metadata/rollout_099/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0023.png", "metadata/rollout_099/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0024.png", "metadata/rollout_099/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0025.png", "metadata/rollout_099/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0026.png", "metadata/rollout_099/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_099/agentview/0027.png", "metadata/rollout_099/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0028.png", "metadata/rollout_099/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0029.png", "metadata/rollout_099/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0030.png", "metadata/rollout_099/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0031.png", "metadata/rollout_099/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0032.png", "metadata/rollout_099/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0033.png", "metadata/rollout_099/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0034.png", "metadata/rollout_099/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0035.png", "metadata/rollout_099/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0036.png", "metadata/rollout_099/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0037.png", "metadata/rollout_099/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0038.png", "metadata/rollout_099/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0039.png", "metadata/rollout_099/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0040.png", "metadata/rollout_099/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0041.png", "metadata/rollout_099/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0042.png", "metadata/rollout_099/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0043.png", "metadata/rollout_099/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0044.png", "metadata/rollout_099/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0045.png", "metadata/rollout_099/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_099/agentview/0046.png", "metadata/rollout_099/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_099/agentview/0047.png", "metadata/rollout_099/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_099/agentview/0048.png", "metadata/rollout_099/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_099/agentview/0049.png", "metadata/rollout_099/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0050.png", "metadata/rollout_099/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_099/agentview/0051.png", "metadata/rollout_099/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the soup can in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_099/agentview/0051.png", "metadata/rollout_099/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0000.png", "metadata/rollout_100/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0001.png", "metadata/rollout_100/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0002.png", "metadata/rollout_100/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0003.png", "metadata/rollout_100/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0004.png", "metadata/rollout_100/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0005.png", "metadata/rollout_100/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0006.png", "metadata/rollout_100/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0007.png", "metadata/rollout_100/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0008.png", "metadata/rollout_100/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0009.png", "metadata/rollout_100/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0010.png", "metadata/rollout_100/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0011.png", "metadata/rollout_100/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0012.png", "metadata/rollout_100/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0013.png", "metadata/rollout_100/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0014.png", "metadata/rollout_100/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0015.png", "metadata/rollout_100/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0016.png", "metadata/rollout_100/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0017.png", "metadata/rollout_100/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0018.png", "metadata/rollout_100/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0019.png", "metadata/rollout_100/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0020.png", "metadata/rollout_100/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0021.png", "metadata/rollout_100/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0022.png", "metadata/rollout_100/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0023.png", "metadata/rollout_100/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0024.png", "metadata/rollout_100/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0025.png", "metadata/rollout_100/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_100/agentview/0026.png", "metadata/rollout_100/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0027.png", "metadata/rollout_100/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0028.png", "metadata/rollout_100/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0029.png", "metadata/rollout_100/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0030.png", "metadata/rollout_100/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0031.png", "metadata/rollout_100/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0032.png", "metadata/rollout_100/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0033.png", "metadata/rollout_100/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0034.png", "metadata/rollout_100/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0035.png", "metadata/rollout_100/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0036.png", "metadata/rollout_100/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0037.png", "metadata/rollout_100/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0038.png", "metadata/rollout_100/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0039.png", "metadata/rollout_100/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0040.png", "metadata/rollout_100/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0041.png", "metadata/rollout_100/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_100/agentview/0042.png", "metadata/rollout_100/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0043.png", "metadata/rollout_100/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0044.png", "metadata/rollout_100/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0045.png", "metadata/rollout_100/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0046.png", "metadata/rollout_100/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0047.png", "metadata/rollout_100/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0048.png", "metadata/rollout_100/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0049.png", "metadata/rollout_100/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0050.png", "metadata/rollout_100/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0051.png", "metadata/rollout_100/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0052.png", "metadata/rollout_100/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0053.png", "metadata/rollout_100/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0054.png", "metadata/rollout_100/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0055.png", "metadata/rollout_100/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0056.png", "metadata/rollout_100/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0057.png", "metadata/rollout_100/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0058.png", "metadata/rollout_100/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0059.png", "metadata/rollout_100/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0060.png", "metadata/rollout_100/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0061.png", "metadata/rollout_100/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0062.png", "metadata/rollout_100/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0063.png", "metadata/rollout_100/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0064.png", "metadata/rollout_100/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0065.png", "metadata/rollout_100/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0066.png", "metadata/rollout_100/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0067.png", "metadata/rollout_100/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0068.png", "metadata/rollout_100/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0069.png", "metadata/rollout_100/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_100/agentview/0070.png", "metadata/rollout_100/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0071.png", "metadata/rollout_100/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0072.png", "metadata/rollout_100/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0073.png", "metadata/rollout_100/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0074.png", "metadata/rollout_100/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0075.png", "metadata/rollout_100/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0076.png", "metadata/rollout_100/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0077.png", "metadata/rollout_100/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0078.png", "metadata/rollout_100/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0079.png", "metadata/rollout_100/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0080.png", "metadata/rollout_100/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0081.png", "metadata/rollout_100/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0082.png", "metadata/rollout_100/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0083.png", "metadata/rollout_100/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0084.png", "metadata/rollout_100/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0085.png", "metadata/rollout_100/wrist/0085.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0086.png", "metadata/rollout_100/wrist/0086.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0087.png", "metadata/rollout_100/wrist/0087.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0088.png", "metadata/rollout_100/wrist/0088.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0089.png", "metadata/rollout_100/wrist/0089.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0090.png", "metadata/rollout_100/wrist/0090.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0091.png", "metadata/rollout_100/wrist/0091.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0092.png", "metadata/rollout_100/wrist/0092.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0093.png", "metadata/rollout_100/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0094.png", "metadata/rollout_100/wrist/0094.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0095.png", "metadata/rollout_100/wrist/0095.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0096.png", "metadata/rollout_100/wrist/0096.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0097.png", "metadata/rollout_100/wrist/0097.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0098.png", "metadata/rollout_100/wrist/0098.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0099.png", "metadata/rollout_100/wrist/0099.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0100.png", "metadata/rollout_100/wrist/0100.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0101.png", "metadata/rollout_100/wrist/0101.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0102.png", "metadata/rollout_100/wrist/0102.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_100/agentview/0103.png", "metadata/rollout_100/wrist/0103.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_100/agentview/0104.png", "metadata/rollout_100/wrist/0104.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0105.png", "metadata/rollout_100/wrist/0105.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_100/agentview/0106.png", "metadata/rollout_100/wrist/0106.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_100/agentview/0107.png", "metadata/rollout_100/wrist/0107.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0108.png", "metadata/rollout_100/wrist/0108.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_100/agentview/0109.png", "metadata/rollout_100/wrist/0109.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_100/agentview/0109.png", "metadata/rollout_100/wrist/0109.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0000.png", "metadata/rollout_101/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0001.png", "metadata/rollout_101/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0002.png", "metadata/rollout_101/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0003.png", "metadata/rollout_101/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0004.png", "metadata/rollout_101/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0005.png", "metadata/rollout_101/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0006.png", "metadata/rollout_101/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0007.png", "metadata/rollout_101/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0008.png", "metadata/rollout_101/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0009.png", "metadata/rollout_101/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0010.png", "metadata/rollout_101/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_101/agentview/0011.png", "metadata/rollout_101/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0012.png", "metadata/rollout_101/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0013.png", "metadata/rollout_101/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0014.png", "metadata/rollout_101/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0015.png", "metadata/rollout_101/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0016.png", "metadata/rollout_101/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0017.png", "metadata/rollout_101/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0018.png", "metadata/rollout_101/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0019.png", "metadata/rollout_101/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0020.png", "metadata/rollout_101/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0021.png", "metadata/rollout_101/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_101/agentview/0022.png", "metadata/rollout_101/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0023.png", "metadata/rollout_101/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0024.png", "metadata/rollout_101/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0025.png", "metadata/rollout_101/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0026.png", "metadata/rollout_101/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0027.png", "metadata/rollout_101/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0028.png", "metadata/rollout_101/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0029.png", "metadata/rollout_101/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0030.png", "metadata/rollout_101/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0031.png", "metadata/rollout_101/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0032.png", "metadata/rollout_101/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0033.png", "metadata/rollout_101/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0034.png", "metadata/rollout_101/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0035.png", "metadata/rollout_101/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0036.png", "metadata/rollout_101/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_101/agentview/0037.png", "metadata/rollout_101/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0038.png", "metadata/rollout_101/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0039.png", "metadata/rollout_101/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0040.png", "metadata/rollout_101/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0041.png", "metadata/rollout_101/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0042.png", "metadata/rollout_101/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0043.png", "metadata/rollout_101/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0044.png", "metadata/rollout_101/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0045.png", "metadata/rollout_101/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0046.png", "metadata/rollout_101/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0047.png", "metadata/rollout_101/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0048.png", "metadata/rollout_101/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0049.png", "metadata/rollout_101/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0050.png", "metadata/rollout_101/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0051.png", "metadata/rollout_101/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0052.png", "metadata/rollout_101/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_101/agentview/0053.png", "metadata/rollout_101/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_101/agentview/0054.png", "metadata/rollout_101/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_101/agentview/0055.png", "metadata/rollout_101/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_101/agentview/0056.png", "metadata/rollout_101/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_101/agentview/0057.png", "metadata/rollout_101/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0058.png", "metadata/rollout_101/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0059.png", "metadata/rollout_101/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0060.png", "metadata/rollout_101/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0061.png", "metadata/rollout_101/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0062.png", "metadata/rollout_101/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0063.png", "metadata/rollout_101/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0064.png", "metadata/rollout_101/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0065.png", "metadata/rollout_101/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0066.png", "metadata/rollout_101/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0067.png", "metadata/rollout_101/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0068.png", "metadata/rollout_101/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0069.png", "metadata/rollout_101/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0070.png", "metadata/rollout_101/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0071.png", "metadata/rollout_101/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0072.png", "metadata/rollout_101/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0073.png", "metadata/rollout_101/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0074.png", "metadata/rollout_101/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0075.png", "metadata/rollout_101/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0076.png", "metadata/rollout_101/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0077.png", "metadata/rollout_101/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0078.png", "metadata/rollout_101/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0079.png", "metadata/rollout_101/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0080.png", "metadata/rollout_101/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0081.png", "metadata/rollout_101/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0082.png", "metadata/rollout_101/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0083.png", "metadata/rollout_101/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0084.png", "metadata/rollout_101/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0085.png", "metadata/rollout_101/wrist/0085.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0086.png", "metadata/rollout_101/wrist/0086.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0087.png", "metadata/rollout_101/wrist/0087.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0088.png", "metadata/rollout_101/wrist/0088.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0089.png", "metadata/rollout_101/wrist/0089.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0090.png", "metadata/rollout_101/wrist/0090.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_101/agentview/0091.png", "metadata/rollout_101/wrist/0091.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0092.png", "metadata/rollout_101/wrist/0092.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_101/agentview/0093.png", "metadata/rollout_101/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_101/agentview/0094.png", "metadata/rollout_101/wrist/0094.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0095.png", "metadata/rollout_101/wrist/0095.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_101/agentview/0096.png", "metadata/rollout_101/wrist/0096.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_101/agentview/0096.png", "metadata/rollout_101/wrist/0096.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0000.png", "metadata/rollout_102/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0001.png", "metadata/rollout_102/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0002.png", "metadata/rollout_102/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0003.png", "metadata/rollout_102/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0004.png", "metadata/rollout_102/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0005.png", "metadata/rollout_102/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0006.png", "metadata/rollout_102/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0007.png", "metadata/rollout_102/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0008.png", "metadata/rollout_102/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0009.png", "metadata/rollout_102/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0010.png", "metadata/rollout_102/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0011.png", "metadata/rollout_102/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_102/agentview/0012.png", "metadata/rollout_102/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0013.png", "metadata/rollout_102/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0014.png", "metadata/rollout_102/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0015.png", "metadata/rollout_102/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0016.png", "metadata/rollout_102/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0017.png", "metadata/rollout_102/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0018.png", "metadata/rollout_102/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0019.png", "metadata/rollout_102/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0020.png", "metadata/rollout_102/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0021.png", "metadata/rollout_102/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0022.png", "metadata/rollout_102/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_102/agentview/0023.png", "metadata/rollout_102/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0024.png", "metadata/rollout_102/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0025.png", "metadata/rollout_102/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0026.png", "metadata/rollout_102/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0027.png", "metadata/rollout_102/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0028.png", "metadata/rollout_102/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0029.png", "metadata/rollout_102/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0030.png", "metadata/rollout_102/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0031.png", "metadata/rollout_102/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0032.png", "metadata/rollout_102/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0033.png", "metadata/rollout_102/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0034.png", "metadata/rollout_102/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0035.png", "metadata/rollout_102/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0036.png", "metadata/rollout_102/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0037.png", "metadata/rollout_102/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_102/agentview/0038.png", "metadata/rollout_102/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0039.png", "metadata/rollout_102/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0040.png", "metadata/rollout_102/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0041.png", "metadata/rollout_102/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0042.png", "metadata/rollout_102/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0043.png", "metadata/rollout_102/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0044.png", "metadata/rollout_102/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0045.png", "metadata/rollout_102/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0046.png", "metadata/rollout_102/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0047.png", "metadata/rollout_102/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0048.png", "metadata/rollout_102/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0049.png", "metadata/rollout_102/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0050.png", "metadata/rollout_102/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0051.png", "metadata/rollout_102/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0052.png", "metadata/rollout_102/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0053.png", "metadata/rollout_102/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0054.png", "metadata/rollout_102/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0055.png", "metadata/rollout_102/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0056.png", "metadata/rollout_102/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0057.png", "metadata/rollout_102/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0058.png", "metadata/rollout_102/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0059.png", "metadata/rollout_102/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0060.png", "metadata/rollout_102/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0061.png", "metadata/rollout_102/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0062.png", "metadata/rollout_102/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0063.png", "metadata/rollout_102/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_102/agentview/0064.png", "metadata/rollout_102/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0065.png", "metadata/rollout_102/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0066.png", "metadata/rollout_102/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0067.png", "metadata/rollout_102/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0068.png", "metadata/rollout_102/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0069.png", "metadata/rollout_102/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0070.png", "metadata/rollout_102/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0071.png", "metadata/rollout_102/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0072.png", "metadata/rollout_102/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0073.png", "metadata/rollout_102/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0074.png", "metadata/rollout_102/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0075.png", "metadata/rollout_102/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0076.png", "metadata/rollout_102/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0077.png", "metadata/rollout_102/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0078.png", "metadata/rollout_102/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0079.png", "metadata/rollout_102/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0080.png", "metadata/rollout_102/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0081.png", "metadata/rollout_102/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0082.png", "metadata/rollout_102/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0083.png", "metadata/rollout_102/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0084.png", "metadata/rollout_102/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0085.png", "metadata/rollout_102/wrist/0085.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0086.png", "metadata/rollout_102/wrist/0086.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0087.png", "metadata/rollout_102/wrist/0087.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0088.png", "metadata/rollout_102/wrist/0088.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0089.png", "metadata/rollout_102/wrist/0089.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0090.png", "metadata/rollout_102/wrist/0090.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0091.png", "metadata/rollout_102/wrist/0091.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0092.png", "metadata/rollout_102/wrist/0092.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0093.png", "metadata/rollout_102/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0094.png", "metadata/rollout_102/wrist/0094.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0095.png", "metadata/rollout_102/wrist/0095.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0096.png", "metadata/rollout_102/wrist/0096.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_102/agentview/0097.png", "metadata/rollout_102/wrist/0097.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0098.png", "metadata/rollout_102/wrist/0098.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_102/agentview/0099.png", "metadata/rollout_102/wrist/0099.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_102/agentview/0100.png", "metadata/rollout_102/wrist/0100.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0101.png", "metadata/rollout_102/wrist/0101.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_102/agentview/0102.png", "metadata/rollout_102/wrist/0102.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_102/agentview/0102.png", "metadata/rollout_102/wrist/0102.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0000.png", "metadata/rollout_103/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0001.png", "metadata/rollout_103/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0002.png", "metadata/rollout_103/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0003.png", "metadata/rollout_103/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0004.png", "metadata/rollout_103/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0005.png", "metadata/rollout_103/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0006.png", "metadata/rollout_103/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0007.png", "metadata/rollout_103/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0008.png", "metadata/rollout_103/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0009.png", "metadata/rollout_103/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0010.png", "metadata/rollout_103/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0011.png", "metadata/rollout_103/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_103/agentview/0012.png", "metadata/rollout_103/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0013.png", "metadata/rollout_103/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0014.png", "metadata/rollout_103/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0015.png", "metadata/rollout_103/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0016.png", "metadata/rollout_103/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0017.png", "metadata/rollout_103/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0018.png", "metadata/rollout_103/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0019.png", "metadata/rollout_103/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0020.png", "metadata/rollout_103/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0021.png", "metadata/rollout_103/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0022.png", "metadata/rollout_103/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0023.png", "metadata/rollout_103/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0024.png", "metadata/rollout_103/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0025.png", "metadata/rollout_103/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0026.png", "metadata/rollout_103/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0027.png", "metadata/rollout_103/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0028.png", "metadata/rollout_103/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0029.png", "metadata/rollout_103/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0030.png", "metadata/rollout_103/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0031.png", "metadata/rollout_103/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0032.png", "metadata/rollout_103/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0033.png", "metadata/rollout_103/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0034.png", "metadata/rollout_103/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0035.png", "metadata/rollout_103/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0036.png", "metadata/rollout_103/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0037.png", "metadata/rollout_103/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_103/agentview/0038.png", "metadata/rollout_103/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0039.png", "metadata/rollout_103/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0040.png", "metadata/rollout_103/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0041.png", "metadata/rollout_103/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0042.png", "metadata/rollout_103/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0043.png", "metadata/rollout_103/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0044.png", "metadata/rollout_103/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0045.png", "metadata/rollout_103/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0046.png", "metadata/rollout_103/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0047.png", "metadata/rollout_103/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0048.png", "metadata/rollout_103/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0049.png", "metadata/rollout_103/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0050.png", "metadata/rollout_103/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0051.png", "metadata/rollout_103/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0052.png", "metadata/rollout_103/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0053.png", "metadata/rollout_103/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0054.png", "metadata/rollout_103/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0055.png", "metadata/rollout_103/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0056.png", "metadata/rollout_103/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0057.png", "metadata/rollout_103/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0058.png", "metadata/rollout_103/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0059.png", "metadata/rollout_103/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0060.png", "metadata/rollout_103/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0061.png", "metadata/rollout_103/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0062.png", "metadata/rollout_103/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0063.png", "metadata/rollout_103/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0064.png", "metadata/rollout_103/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0065.png", "metadata/rollout_103/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0066.png", "metadata/rollout_103/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0067.png", "metadata/rollout_103/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0068.png", "metadata/rollout_103/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0069.png", "metadata/rollout_103/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0070.png", "metadata/rollout_103/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0071.png", "metadata/rollout_103/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0072.png", "metadata/rollout_103/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0073.png", "metadata/rollout_103/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0074.png", "metadata/rollout_103/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0075.png", "metadata/rollout_103/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0076.png", "metadata/rollout_103/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0077.png", "metadata/rollout_103/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0078.png", "metadata/rollout_103/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0079.png", "metadata/rollout_103/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0080.png", "metadata/rollout_103/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0081.png", "metadata/rollout_103/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0082.png", "metadata/rollout_103/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0083.png", "metadata/rollout_103/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_103/agentview/0084.png", "metadata/rollout_103/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_103/agentview/0085.png", "metadata/rollout_103/wrist/0085.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_103/agentview/0086.png", "metadata/rollout_103/wrist/0086.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_103/agentview/0087.png", "metadata/rollout_103/wrist/0087.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_103/agentview/0088.png", "metadata/rollout_103/wrist/0088.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_103/agentview/0089.png", "metadata/rollout_103/wrist/0089.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_103/agentview/0090.png", "metadata/rollout_103/wrist/0090.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_103/agentview/0091.png", "metadata/rollout_103/wrist/0091.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_103/agentview/0092.png", "metadata/rollout_103/wrist/0092.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_103/agentview/0093.png", "metadata/rollout_103/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_103/agentview/0094.png", "metadata/rollout_103/wrist/0094.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_103/agentview/0095.png", "metadata/rollout_103/wrist/0095.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0096.png", "metadata/rollout_103/wrist/0096.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_RIGHT, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_103/agentview/0097.png", "metadata/rollout_103/wrist/0097.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_103/agentview/0097.png", "metadata/rollout_103/wrist/0097.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0000.png", "metadata/rollout_104/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0001.png", "metadata/rollout_104/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0002.png", "metadata/rollout_104/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0003.png", "metadata/rollout_104/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0004.png", "metadata/rollout_104/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0005.png", "metadata/rollout_104/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0006.png", "metadata/rollout_104/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0007.png", "metadata/rollout_104/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0008.png", "metadata/rollout_104/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0009.png", "metadata/rollout_104/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0010.png", "metadata/rollout_104/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0011.png", "metadata/rollout_104/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0012.png", "metadata/rollout_104/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0013.png", "metadata/rollout_104/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0014.png", "metadata/rollout_104/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0015.png", "metadata/rollout_104/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0016.png", "metadata/rollout_104/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0017.png", "metadata/rollout_104/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0018.png", "metadata/rollout_104/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0019.png", "metadata/rollout_104/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0020.png", "metadata/rollout_104/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0021.png", "metadata/rollout_104/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0022.png", "metadata/rollout_104/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0023.png", "metadata/rollout_104/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0024.png", "metadata/rollout_104/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_104/agentview/0025.png", "metadata/rollout_104/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0026.png", "metadata/rollout_104/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0027.png", "metadata/rollout_104/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0028.png", "metadata/rollout_104/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0029.png", "metadata/rollout_104/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0030.png", "metadata/rollout_104/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0031.png", "metadata/rollout_104/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0032.png", "metadata/rollout_104/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0033.png", "metadata/rollout_104/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0034.png", "metadata/rollout_104/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0035.png", "metadata/rollout_104/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0036.png", "metadata/rollout_104/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0037.png", "metadata/rollout_104/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0038.png", "metadata/rollout_104/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0039.png", "metadata/rollout_104/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0040.png", "metadata/rollout_104/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_104/agentview/0041.png", "metadata/rollout_104/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0042.png", "metadata/rollout_104/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0043.png", "metadata/rollout_104/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0044.png", "metadata/rollout_104/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0045.png", "metadata/rollout_104/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0046.png", "metadata/rollout_104/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0047.png", "metadata/rollout_104/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0048.png", "metadata/rollout_104/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0049.png", "metadata/rollout_104/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0050.png", "metadata/rollout_104/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0051.png", "metadata/rollout_104/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0052.png", "metadata/rollout_104/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0053.png", "metadata/rollout_104/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0054.png", "metadata/rollout_104/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0055.png", "metadata/rollout_104/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0056.png", "metadata/rollout_104/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0057.png", "metadata/rollout_104/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0058.png", "metadata/rollout_104/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0059.png", "metadata/rollout_104/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0060.png", "metadata/rollout_104/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0061.png", "metadata/rollout_104/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0062.png", "metadata/rollout_104/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0063.png", "metadata/rollout_104/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0064.png", "metadata/rollout_104/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0065.png", "metadata/rollout_104/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0066.png", "metadata/rollout_104/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0067.png", "metadata/rollout_104/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0068.png", "metadata/rollout_104/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0069.png", "metadata/rollout_104/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0070.png", "metadata/rollout_104/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0071.png", "metadata/rollout_104/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0072.png", "metadata/rollout_104/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0073.png", "metadata/rollout_104/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0074.png", "metadata/rollout_104/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0075.png", "metadata/rollout_104/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0076.png", "metadata/rollout_104/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0077.png", "metadata/rollout_104/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0078.png", "metadata/rollout_104/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0079.png", "metadata/rollout_104/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0080.png", "metadata/rollout_104/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0081.png", "metadata/rollout_104/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0082.png", "metadata/rollout_104/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0083.png", "metadata/rollout_104/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0084.png", "metadata/rollout_104/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0085.png", "metadata/rollout_104/wrist/0085.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0086.png", "metadata/rollout_104/wrist/0086.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0087.png", "metadata/rollout_104/wrist/0087.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0088.png", "metadata/rollout_104/wrist/0088.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0089.png", "metadata/rollout_104/wrist/0089.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_104/agentview/0090.png", "metadata/rollout_104/wrist/0090.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0091.png", "metadata/rollout_104/wrist/0091.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0092.png", "metadata/rollout_104/wrist/0092.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0093.png", "metadata/rollout_104/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0094.png", "metadata/rollout_104/wrist/0094.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0095.png", "metadata/rollout_104/wrist/0095.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0096.png", "metadata/rollout_104/wrist/0096.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0097.png", "metadata/rollout_104/wrist/0097.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0098.png", "metadata/rollout_104/wrist/0098.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0099.png", "metadata/rollout_104/wrist/0099.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_104/agentview/0100.png", "metadata/rollout_104/wrist/0100.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_104/agentview/0101.png", "metadata/rollout_104/wrist/0101.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_104/agentview/0102.png", "metadata/rollout_104/wrist/0102.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_104/agentview/0103.png", "metadata/rollout_104/wrist/0103.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0104.png", "metadata/rollout_104/wrist/0104.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_104/agentview/0105.png", "metadata/rollout_104/wrist/0105.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_104/agentview/0105.png", "metadata/rollout_104/wrist/0105.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0000.png", "metadata/rollout_105/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0001.png", "metadata/rollout_105/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0002.png", "metadata/rollout_105/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0003.png", "metadata/rollout_105/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0004.png", "metadata/rollout_105/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0005.png", "metadata/rollout_105/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0006.png", "metadata/rollout_105/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0007.png", "metadata/rollout_105/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0008.png", "metadata/rollout_105/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0009.png", "metadata/rollout_105/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0010.png", "metadata/rollout_105/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0011.png", "metadata/rollout_105/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0012.png", "metadata/rollout_105/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0013.png", "metadata/rollout_105/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0014.png", "metadata/rollout_105/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0015.png", "metadata/rollout_105/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0016.png", "metadata/rollout_105/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0017.png", "metadata/rollout_105/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0018.png", "metadata/rollout_105/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0019.png", "metadata/rollout_105/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0020.png", "metadata/rollout_105/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0021.png", "metadata/rollout_105/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0022.png", "metadata/rollout_105/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0023.png", "metadata/rollout_105/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0024.png", "metadata/rollout_105/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0025.png", "metadata/rollout_105/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0026.png", "metadata/rollout_105/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0027.png", "metadata/rollout_105/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0028.png", "metadata/rollout_105/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0029.png", "metadata/rollout_105/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0030.png", "metadata/rollout_105/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0031.png", "metadata/rollout_105/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0032.png", "metadata/rollout_105/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_105/agentview/0033.png", "metadata/rollout_105/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0034.png", "metadata/rollout_105/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0035.png", "metadata/rollout_105/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0036.png", "metadata/rollout_105/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0037.png", "metadata/rollout_105/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0038.png", "metadata/rollout_105/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0039.png", "metadata/rollout_105/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0040.png", "metadata/rollout_105/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_105/agentview/0041.png", "metadata/rollout_105/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0042.png", "metadata/rollout_105/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0043.png", "metadata/rollout_105/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0044.png", "metadata/rollout_105/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0045.png", "metadata/rollout_105/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0046.png", "metadata/rollout_105/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0047.png", "metadata/rollout_105/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0048.png", "metadata/rollout_105/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0049.png", "metadata/rollout_105/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0050.png", "metadata/rollout_105/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0051.png", "metadata/rollout_105/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0052.png", "metadata/rollout_105/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0053.png", "metadata/rollout_105/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0054.png", "metadata/rollout_105/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0055.png", "metadata/rollout_105/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0056.png", "metadata/rollout_105/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0057.png", "metadata/rollout_105/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0058.png", "metadata/rollout_105/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0059.png", "metadata/rollout_105/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0060.png", "metadata/rollout_105/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0061.png", "metadata/rollout_105/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0062.png", "metadata/rollout_105/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0063.png", "metadata/rollout_105/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0064.png", "metadata/rollout_105/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0065.png", "metadata/rollout_105/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0066.png", "metadata/rollout_105/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0067.png", "metadata/rollout_105/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0068.png", "metadata/rollout_105/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0069.png", "metadata/rollout_105/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0070.png", "metadata/rollout_105/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0071.png", "metadata/rollout_105/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0072.png", "metadata/rollout_105/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0073.png", "metadata/rollout_105/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0074.png", "metadata/rollout_105/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0075.png", "metadata/rollout_105/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0076.png", "metadata/rollout_105/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0077.png", "metadata/rollout_105/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0078.png", "metadata/rollout_105/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0079.png", "metadata/rollout_105/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0080.png", "metadata/rollout_105/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0081.png", "metadata/rollout_105/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0082.png", "metadata/rollout_105/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0083.png", "metadata/rollout_105/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0084.png", "metadata/rollout_105/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0085.png", "metadata/rollout_105/wrist/0085.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0086.png", "metadata/rollout_105/wrist/0086.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0087.png", "metadata/rollout_105/wrist/0087.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0088.png", "metadata/rollout_105/wrist/0088.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0089.png", "metadata/rollout_105/wrist/0089.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0090.png", "metadata/rollout_105/wrist/0090.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0091.png", "metadata/rollout_105/wrist/0091.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_105/agentview/0092.png", "metadata/rollout_105/wrist/0092.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_105/agentview/0093.png", "metadata/rollout_105/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_105/agentview/0094.png", "metadata/rollout_105/wrist/0094.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_105/agentview/0095.png", "metadata/rollout_105/wrist/0095.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_105/agentview/0096.png", "metadata/rollout_105/wrist/0096.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_105/agentview/0097.png", "metadata/rollout_105/wrist/0097.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_105/agentview/0098.png", "metadata/rollout_105/wrist/0098.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_105/agentview/0099.png", "metadata/rollout_105/wrist/0099.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_105/agentview/0100.png", "metadata/rollout_105/wrist/0100.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_105/agentview/0101.png", "metadata/rollout_105/wrist/0101.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_105/agentview/0102.png", "metadata/rollout_105/wrist/0102.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_105/agentview/0103.png", "metadata/rollout_105/wrist/0103.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0104.png", "metadata/rollout_105/wrist/0104.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_105/agentview/0105.png", "metadata/rollout_105/wrist/0105.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_105/agentview/0105.png", "metadata/rollout_105/wrist/0105.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0000.png", "metadata/rollout_106/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0001.png", "metadata/rollout_106/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0002.png", "metadata/rollout_106/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0003.png", "metadata/rollout_106/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0004.png", "metadata/rollout_106/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0005.png", "metadata/rollout_106/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0006.png", "metadata/rollout_106/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0007.png", "metadata/rollout_106/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0008.png", "metadata/rollout_106/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0009.png", "metadata/rollout_106/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0010.png", "metadata/rollout_106/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0011.png", "metadata/rollout_106/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_106/agentview/0012.png", "metadata/rollout_106/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_106/agentview/0013.png", "metadata/rollout_106/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_106/agentview/0014.png", "metadata/rollout_106/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_106/agentview/0015.png", "metadata/rollout_106/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_106/agentview/0016.png", "metadata/rollout_106/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_106/agentview/0017.png", "metadata/rollout_106/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_106/agentview/0018.png", "metadata/rollout_106/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_106/agentview/0019.png", "metadata/rollout_106/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_106/agentview/0020.png", "metadata/rollout_106/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_106/agentview/0021.png", "metadata/rollout_106/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0022.png", "metadata/rollout_106/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0023.png", "metadata/rollout_106/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0024.png", "metadata/rollout_106/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0025.png", "metadata/rollout_106/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0026.png", "metadata/rollout_106/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0027.png", "metadata/rollout_106/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0028.png", "metadata/rollout_106/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0029.png", "metadata/rollout_106/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0030.png", "metadata/rollout_106/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0031.png", "metadata/rollout_106/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0032.png", "metadata/rollout_106/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0033.png", "metadata/rollout_106/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0034.png", "metadata/rollout_106/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0035.png", "metadata/rollout_106/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_106/agentview/0036.png", "metadata/rollout_106/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0037.png", "metadata/rollout_106/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0038.png", "metadata/rollout_106/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0039.png", "metadata/rollout_106/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0040.png", "metadata/rollout_106/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0041.png", "metadata/rollout_106/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0042.png", "metadata/rollout_106/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0043.png", "metadata/rollout_106/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0044.png", "metadata/rollout_106/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0045.png", "metadata/rollout_106/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0046.png", "metadata/rollout_106/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0047.png", "metadata/rollout_106/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0048.png", "metadata/rollout_106/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0049.png", "metadata/rollout_106/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0050.png", "metadata/rollout_106/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0051.png", "metadata/rollout_106/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0052.png", "metadata/rollout_106/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0053.png", "metadata/rollout_106/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0054.png", "metadata/rollout_106/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0055.png", "metadata/rollout_106/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0056.png", "metadata/rollout_106/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0057.png", "metadata/rollout_106/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0058.png", "metadata/rollout_106/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0059.png", "metadata/rollout_106/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0060.png", "metadata/rollout_106/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0061.png", "metadata/rollout_106/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0062.png", "metadata/rollout_106/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0063.png", "metadata/rollout_106/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0064.png", "metadata/rollout_106/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0065.png", "metadata/rollout_106/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0066.png", "metadata/rollout_106/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0067.png", "metadata/rollout_106/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0068.png", "metadata/rollout_106/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0069.png", "metadata/rollout_106/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0070.png", "metadata/rollout_106/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0071.png", "metadata/rollout_106/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0072.png", "metadata/rollout_106/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0073.png", "metadata/rollout_106/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0074.png", "metadata/rollout_106/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0075.png", "metadata/rollout_106/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0076.png", "metadata/rollout_106/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0077.png", "metadata/rollout_106/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0078.png", "metadata/rollout_106/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0079.png", "metadata/rollout_106/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0080.png", "metadata/rollout_106/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0081.png", "metadata/rollout_106/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_106/agentview/0082.png", "metadata/rollout_106/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_106/agentview/0083.png", "metadata/rollout_106/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_106/agentview/0084.png", "metadata/rollout_106/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_106/agentview/0085.png", "metadata/rollout_106/wrist/0085.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_106/agentview/0086.png", "metadata/rollout_106/wrist/0086.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_106/agentview/0087.png", "metadata/rollout_106/wrist/0087.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_106/agentview/0088.png", "metadata/rollout_106/wrist/0088.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_106/agentview/0089.png", "metadata/rollout_106/wrist/0089.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_106/agentview/0090.png", "metadata/rollout_106/wrist/0090.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_106/agentview/0091.png", "metadata/rollout_106/wrist/0091.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0092.png", "metadata/rollout_106/wrist/0092.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_106/agentview/0093.png", "metadata/rollout_106/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_106/agentview/0093.png", "metadata/rollout_106/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0000.png", "metadata/rollout_107/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0001.png", "metadata/rollout_107/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0002.png", "metadata/rollout_107/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0003.png", "metadata/rollout_107/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0004.png", "metadata/rollout_107/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0005.png", "metadata/rollout_107/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0006.png", "metadata/rollout_107/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0007.png", "metadata/rollout_107/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0008.png", "metadata/rollout_107/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_107/agentview/0009.png", "metadata/rollout_107/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_107/agentview/0010.png", "metadata/rollout_107/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_107/agentview/0011.png", "metadata/rollout_107/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_107/agentview/0012.png", "metadata/rollout_107/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_107/agentview/0013.png", "metadata/rollout_107/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_107/agentview/0014.png", "metadata/rollout_107/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_107/agentview/0015.png", "metadata/rollout_107/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_107/agentview/0016.png", "metadata/rollout_107/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_107/agentview/0017.png", "metadata/rollout_107/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_107/agentview/0018.png", "metadata/rollout_107/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_107/agentview/0019.png", "metadata/rollout_107/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_107/agentview/0020.png", "metadata/rollout_107/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_107/agentview/0021.png", "metadata/rollout_107/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_107/agentview/0022.png", "metadata/rollout_107/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_107/agentview/0023.png", "metadata/rollout_107/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_107/agentview/0024.png", "metadata/rollout_107/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_107/agentview/0025.png", "metadata/rollout_107/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_107/agentview/0026.png", "metadata/rollout_107/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0027.png", "metadata/rollout_107/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0028.png", "metadata/rollout_107/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0029.png", "metadata/rollout_107/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0030.png", "metadata/rollout_107/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0031.png", "metadata/rollout_107/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_107/agentview/0032.png", "metadata/rollout_107/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0033.png", "metadata/rollout_107/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0034.png", "metadata/rollout_107/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0035.png", "metadata/rollout_107/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0036.png", "metadata/rollout_107/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0037.png", "metadata/rollout_107/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0038.png", "metadata/rollout_107/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0039.png", "metadata/rollout_107/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0040.png", "metadata/rollout_107/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0041.png", "metadata/rollout_107/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0042.png", "metadata/rollout_107/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0043.png", "metadata/rollout_107/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0044.png", "metadata/rollout_107/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0045.png", "metadata/rollout_107/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0046.png", "metadata/rollout_107/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0047.png", "metadata/rollout_107/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0048.png", "metadata/rollout_107/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0049.png", "metadata/rollout_107/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0050.png", "metadata/rollout_107/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0051.png", "metadata/rollout_107/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0052.png", "metadata/rollout_107/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0053.png", "metadata/rollout_107/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0054.png", "metadata/rollout_107/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0055.png", "metadata/rollout_107/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0056.png", "metadata/rollout_107/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0057.png", "metadata/rollout_107/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0058.png", "metadata/rollout_107/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0059.png", "metadata/rollout_107/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0060.png", "metadata/rollout_107/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0061.png", "metadata/rollout_107/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0062.png", "metadata/rollout_107/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0063.png", "metadata/rollout_107/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0064.png", "metadata/rollout_107/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0065.png", "metadata/rollout_107/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0066.png", "metadata/rollout_107/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0067.png", "metadata/rollout_107/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0068.png", "metadata/rollout_107/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0069.png", "metadata/rollout_107/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0070.png", "metadata/rollout_107/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0071.png", "metadata/rollout_107/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_107/agentview/0072.png", "metadata/rollout_107/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_107/agentview/0073.png", "metadata/rollout_107/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_107/agentview/0074.png", "metadata/rollout_107/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_107/agentview/0075.png", "metadata/rollout_107/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_107/agentview/0076.png", "metadata/rollout_107/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_107/agentview/0077.png", "metadata/rollout_107/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_107/agentview/0078.png", "metadata/rollout_107/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_107/agentview/0079.png", "metadata/rollout_107/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0080.png", "metadata/rollout_107/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_107/agentview/0081.png", "metadata/rollout_107/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_107/agentview/0082.png", "metadata/rollout_107/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0083.png", "metadata/rollout_107/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_107/agentview/0084.png", "metadata/rollout_107/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_107/agentview/0084.png", "metadata/rollout_107/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0000.png", "metadata/rollout_108/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0001.png", "metadata/rollout_108/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0002.png", "metadata/rollout_108/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0003.png", "metadata/rollout_108/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0004.png", "metadata/rollout_108/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0005.png", "metadata/rollout_108/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0006.png", "metadata/rollout_108/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0007.png", "metadata/rollout_108/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0008.png", "metadata/rollout_108/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0009.png", "metadata/rollout_108/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0010.png", "metadata/rollout_108/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0011.png", "metadata/rollout_108/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0012.png", "metadata/rollout_108/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0013.png", "metadata/rollout_108/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0014.png", "metadata/rollout_108/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0015.png", "metadata/rollout_108/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0016.png", "metadata/rollout_108/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0017.png", "metadata/rollout_108/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0018.png", "metadata/rollout_108/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0019.png", "metadata/rollout_108/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0020.png", "metadata/rollout_108/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0021.png", "metadata/rollout_108/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0022.png", "metadata/rollout_108/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0023.png", "metadata/rollout_108/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0024.png", "metadata/rollout_108/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0025.png", "metadata/rollout_108/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0026.png", "metadata/rollout_108/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0027.png", "metadata/rollout_108/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0028.png", "metadata/rollout_108/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0029.png", "metadata/rollout_108/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0030.png", "metadata/rollout_108/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0031.png", "metadata/rollout_108/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0032.png", "metadata/rollout_108/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0033.png", "metadata/rollout_108/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0034.png", "metadata/rollout_108/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0035.png", "metadata/rollout_108/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0036.png", "metadata/rollout_108/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0037.png", "metadata/rollout_108/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0038.png", "metadata/rollout_108/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0039.png", "metadata/rollout_108/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0040.png", "metadata/rollout_108/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0041.png", "metadata/rollout_108/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0042.png", "metadata/rollout_108/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0043.png", "metadata/rollout_108/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_108/agentview/0044.png", "metadata/rollout_108/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0045.png", "metadata/rollout_108/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0046.png", "metadata/rollout_108/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0047.png", "metadata/rollout_108/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0048.png", "metadata/rollout_108/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0049.png", "metadata/rollout_108/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0050.png", "metadata/rollout_108/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0051.png", "metadata/rollout_108/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0052.png", "metadata/rollout_108/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0053.png", "metadata/rollout_108/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0054.png", "metadata/rollout_108/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0055.png", "metadata/rollout_108/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0056.png", "metadata/rollout_108/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0057.png", "metadata/rollout_108/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0058.png", "metadata/rollout_108/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0059.png", "metadata/rollout_108/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0060.png", "metadata/rollout_108/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0061.png", "metadata/rollout_108/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0062.png", "metadata/rollout_108/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0063.png", "metadata/rollout_108/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0064.png", "metadata/rollout_108/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0065.png", "metadata/rollout_108/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0066.png", "metadata/rollout_108/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0067.png", "metadata/rollout_108/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0068.png", "metadata/rollout_108/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0069.png", "metadata/rollout_108/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0070.png", "metadata/rollout_108/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0071.png", "metadata/rollout_108/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0072.png", "metadata/rollout_108/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0073.png", "metadata/rollout_108/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0074.png", "metadata/rollout_108/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0075.png", "metadata/rollout_108/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0076.png", "metadata/rollout_108/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0077.png", "metadata/rollout_108/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0078.png", "metadata/rollout_108/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0079.png", "metadata/rollout_108/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0080.png", "metadata/rollout_108/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0081.png", "metadata/rollout_108/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0082.png", "metadata/rollout_108/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0083.png", "metadata/rollout_108/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0084.png", "metadata/rollout_108/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0085.png", "metadata/rollout_108/wrist/0085.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0086.png", "metadata/rollout_108/wrist/0086.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0087.png", "metadata/rollout_108/wrist/0087.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0088.png", "metadata/rollout_108/wrist/0088.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0089.png", "metadata/rollout_108/wrist/0089.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0090.png", "metadata/rollout_108/wrist/0090.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0091.png", "metadata/rollout_108/wrist/0091.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_108/agentview/0092.png", "metadata/rollout_108/wrist/0092.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_108/agentview/0093.png", "metadata/rollout_108/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_108/agentview/0094.png", "metadata/rollout_108/wrist/0094.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_108/agentview/0095.png", "metadata/rollout_108/wrist/0095.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_108/agentview/0096.png", "metadata/rollout_108/wrist/0096.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_108/agentview/0097.png", "metadata/rollout_108/wrist/0097.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_108/agentview/0098.png", "metadata/rollout_108/wrist/0098.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_108/agentview/0099.png", "metadata/rollout_108/wrist/0099.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_108/agentview/0100.png", "metadata/rollout_108/wrist/0100.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_108/agentview/0101.png", "metadata/rollout_108/wrist/0101.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_108/agentview/0102.png", "metadata/rollout_108/wrist/0102.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_108/agentview/0103.png", "metadata/rollout_108/wrist/0103.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_108/agentview/0104.png", "metadata/rollout_108/wrist/0104.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_108/agentview/0105.png", "metadata/rollout_108/wrist/0105.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0106.png", "metadata/rollout_108/wrist/0106.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_LEFT, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_108/agentview/0107.png", "metadata/rollout_108/wrist/0107.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_108/agentview/0107.png", "metadata/rollout_108/wrist/0107.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0000.png", "metadata/rollout_109/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0001.png", "metadata/rollout_109/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0002.png", "metadata/rollout_109/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0003.png", "metadata/rollout_109/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0004.png", "metadata/rollout_109/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0005.png", "metadata/rollout_109/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0006.png", "metadata/rollout_109/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0007.png", "metadata/rollout_109/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0008.png", "metadata/rollout_109/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0009.png", "metadata/rollout_109/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0010.png", "metadata/rollout_109/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0011.png", "metadata/rollout_109/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0012.png", "metadata/rollout_109/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0013.png", "metadata/rollout_109/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0014.png", "metadata/rollout_109/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_109/agentview/0015.png", "metadata/rollout_109/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0016.png", "metadata/rollout_109/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0017.png", "metadata/rollout_109/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0018.png", "metadata/rollout_109/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0019.png", "metadata/rollout_109/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0020.png", "metadata/rollout_109/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0021.png", "metadata/rollout_109/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0022.png", "metadata/rollout_109/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0023.png", "metadata/rollout_109/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0024.png", "metadata/rollout_109/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0025.png", "metadata/rollout_109/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0026.png", "metadata/rollout_109/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0027.png", "metadata/rollout_109/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_109/agentview/0028.png", "metadata/rollout_109/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0029.png", "metadata/rollout_109/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0030.png", "metadata/rollout_109/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0031.png", "metadata/rollout_109/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0032.png", "metadata/rollout_109/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0033.png", "metadata/rollout_109/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0034.png", "metadata/rollout_109/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0035.png", "metadata/rollout_109/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0036.png", "metadata/rollout_109/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0037.png", "metadata/rollout_109/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0038.png", "metadata/rollout_109/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0039.png", "metadata/rollout_109/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0040.png", "metadata/rollout_109/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0041.png", "metadata/rollout_109/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0042.png", "metadata/rollout_109/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_109/agentview/0043.png", "metadata/rollout_109/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0044.png", "metadata/rollout_109/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0045.png", "metadata/rollout_109/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0046.png", "metadata/rollout_109/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0047.png", "metadata/rollout_109/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0048.png", "metadata/rollout_109/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0049.png", "metadata/rollout_109/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0050.png", "metadata/rollout_109/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0051.png", "metadata/rollout_109/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0052.png", "metadata/rollout_109/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0053.png", "metadata/rollout_109/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0054.png", "metadata/rollout_109/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0055.png", "metadata/rollout_109/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0056.png", "metadata/rollout_109/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0057.png", "metadata/rollout_109/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0058.png", "metadata/rollout_109/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0059.png", "metadata/rollout_109/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0060.png", "metadata/rollout_109/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0061.png", "metadata/rollout_109/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0062.png", "metadata/rollout_109/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0063.png", "metadata/rollout_109/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0064.png", "metadata/rollout_109/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0065.png", "metadata/rollout_109/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0066.png", "metadata/rollout_109/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0067.png", "metadata/rollout_109/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0068.png", "metadata/rollout_109/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0069.png", "metadata/rollout_109/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0070.png", "metadata/rollout_109/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0071.png", "metadata/rollout_109/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0072.png", "metadata/rollout_109/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0073.png", "metadata/rollout_109/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_109/agentview/0074.png", "metadata/rollout_109/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0075.png", "metadata/rollout_109/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0076.png", "metadata/rollout_109/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0077.png", "metadata/rollout_109/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0078.png", "metadata/rollout_109/wrist/0078.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0079.png", "metadata/rollout_109/wrist/0079.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0080.png", "metadata/rollout_109/wrist/0080.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0081.png", "metadata/rollout_109/wrist/0081.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0082.png", "metadata/rollout_109/wrist/0082.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0083.png", "metadata/rollout_109/wrist/0083.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0084.png", "metadata/rollout_109/wrist/0084.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0085.png", "metadata/rollout_109/wrist/0085.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0086.png", "metadata/rollout_109/wrist/0086.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0087.png", "metadata/rollout_109/wrist/0087.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0088.png", "metadata/rollout_109/wrist/0088.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0089.png", "metadata/rollout_109/wrist/0089.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0090.png", "metadata/rollout_109/wrist/0090.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0091.png", "metadata/rollout_109/wrist/0091.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0092.png", "metadata/rollout_109/wrist/0092.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0093.png", "metadata/rollout_109/wrist/0093.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0094.png", "metadata/rollout_109/wrist/0094.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0095.png", "metadata/rollout_109/wrist/0095.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0096.png", "metadata/rollout_109/wrist/0096.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0097.png", "metadata/rollout_109/wrist/0097.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0098.png", "metadata/rollout_109/wrist/0098.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0099.png", "metadata/rollout_109/wrist/0099.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0100.png", "metadata/rollout_109/wrist/0100.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0101.png", "metadata/rollout_109/wrist/0101.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0102.png", "metadata/rollout_109/wrist/0102.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0103.png", "metadata/rollout_109/wrist/0103.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0104.png", "metadata/rollout_109/wrist/0104.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_109/agentview/0105.png", "metadata/rollout_109/wrist/0105.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0106.png", "metadata/rollout_109/wrist/0106.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_109/agentview/0107.png", "metadata/rollout_109/wrist/0107.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_109/agentview/0108.png", "metadata/rollout_109/wrist/0108.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0109.png", "metadata/rollout_109/wrist/0109.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_109/agentview/0110.png", "metadata/rollout_109/wrist/0110.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Throw away the apple\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_109/agentview/0110.png", "metadata/rollout_109/wrist/0110.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0000.png", "metadata/rollout_110/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0001.png", "metadata/rollout_110/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0002.png", "metadata/rollout_110/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0003.png", "metadata/rollout_110/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0004.png", "metadata/rollout_110/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0005.png", "metadata/rollout_110/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0006.png", "metadata/rollout_110/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0007.png", "metadata/rollout_110/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0008.png", "metadata/rollout_110/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_110/agentview/0009.png", "metadata/rollout_110/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_110/agentview/0010.png", "metadata/rollout_110/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_110/agentview/0011.png", "metadata/rollout_110/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_110/agentview/0012.png", "metadata/rollout_110/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_110/agentview/0013.png", "metadata/rollout_110/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_110/agentview/0014.png", "metadata/rollout_110/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_110/agentview/0015.png", "metadata/rollout_110/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0016.png", "metadata/rollout_110/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0017.png", "metadata/rollout_110/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0018.png", "metadata/rollout_110/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0019.png", "metadata/rollout_110/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0020.png", "metadata/rollout_110/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0021.png", "metadata/rollout_110/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_110/agentview/0022.png", "metadata/rollout_110/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0023.png", "metadata/rollout_110/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0024.png", "metadata/rollout_110/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0025.png", "metadata/rollout_110/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0026.png", "metadata/rollout_110/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0027.png", "metadata/rollout_110/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0028.png", "metadata/rollout_110/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0029.png", "metadata/rollout_110/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0030.png", "metadata/rollout_110/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_110/agentview/0031.png", "metadata/rollout_110/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_110/agentview/0032.png", "metadata/rollout_110/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_110/agentview/0033.png", "metadata/rollout_110/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_110/agentview/0034.png", "metadata/rollout_110/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_110/agentview/0035.png", "metadata/rollout_110/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_110/agentview/0036.png", "metadata/rollout_110/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_110/agentview/0037.png", "metadata/rollout_110/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_110/agentview/0038.png", "metadata/rollout_110/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_110/agentview/0039.png", "metadata/rollout_110/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0040.png", "metadata/rollout_110/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0041.png", "metadata/rollout_110/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0042.png", "metadata/rollout_110/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_110/agentview/0043.png", "metadata/rollout_110/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_110/agentview/0044.png", "metadata/rollout_110/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0045.png", "metadata/rollout_110/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_110/agentview/0046.png", "metadata/rollout_110/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_110/agentview/0046.png", "metadata/rollout_110/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0000.png", "metadata/rollout_111/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0001.png", "metadata/rollout_111/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0002.png", "metadata/rollout_111/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0003.png", "metadata/rollout_111/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0004.png", "metadata/rollout_111/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0005.png", "metadata/rollout_111/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0006.png", "metadata/rollout_111/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0007.png", "metadata/rollout_111/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0008.png", "metadata/rollout_111/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_111/agentview/0009.png", "metadata/rollout_111/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_111/agentview/0010.png", "metadata/rollout_111/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_111/agentview/0011.png", "metadata/rollout_111/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_111/agentview/0012.png", "metadata/rollout_111/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_111/agentview/0013.png", "metadata/rollout_111/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_111/agentview/0014.png", "metadata/rollout_111/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_111/agentview/0015.png", "metadata/rollout_111/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_111/agentview/0016.png", "metadata/rollout_111/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_111/agentview/0017.png", "metadata/rollout_111/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0018.png", "metadata/rollout_111/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0019.png", "metadata/rollout_111/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0020.png", "metadata/rollout_111/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0021.png", "metadata/rollout_111/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0022.png", "metadata/rollout_111/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0023.png", "metadata/rollout_111/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_111/agentview/0024.png", "metadata/rollout_111/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0025.png", "metadata/rollout_111/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0026.png", "metadata/rollout_111/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0027.png", "metadata/rollout_111/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0028.png", "metadata/rollout_111/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0029.png", "metadata/rollout_111/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0030.png", "metadata/rollout_111/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0031.png", "metadata/rollout_111/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0032.png", "metadata/rollout_111/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0033.png", "metadata/rollout_111/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0034.png", "metadata/rollout_111/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0035.png", "metadata/rollout_111/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0036.png", "metadata/rollout_111/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0037.png", "metadata/rollout_111/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0038.png", "metadata/rollout_111/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0039.png", "metadata/rollout_111/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0040.png", "metadata/rollout_111/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0041.png", "metadata/rollout_111/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0042.png", "metadata/rollout_111/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_111/agentview/0043.png", "metadata/rollout_111/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_111/agentview/0044.png", "metadata/rollout_111/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_111/agentview/0045.png", "metadata/rollout_111/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_111/agentview/0046.png", "metadata/rollout_111/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0047.png", "metadata/rollout_111/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0048.png", "metadata/rollout_111/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0049.png", "metadata/rollout_111/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_111/agentview/0050.png", "metadata/rollout_111/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_111/agentview/0051.png", "metadata/rollout_111/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0052.png", "metadata/rollout_111/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_111/agentview/0053.png", "metadata/rollout_111/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_111/agentview/0053.png", "metadata/rollout_111/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0000.png", "metadata/rollout_112/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0001.png", "metadata/rollout_112/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0002.png", "metadata/rollout_112/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0003.png", "metadata/rollout_112/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0004.png", "metadata/rollout_112/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0005.png", "metadata/rollout_112/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0006.png", "metadata/rollout_112/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0007.png", "metadata/rollout_112/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0008.png", "metadata/rollout_112/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_112/agentview/0009.png", "metadata/rollout_112/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_112/agentview/0010.png", "metadata/rollout_112/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_112/agentview/0011.png", "metadata/rollout_112/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_112/agentview/0012.png", "metadata/rollout_112/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_112/agentview/0013.png", "metadata/rollout_112/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_112/agentview/0014.png", "metadata/rollout_112/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_112/agentview/0015.png", "metadata/rollout_112/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_112/agentview/0016.png", "metadata/rollout_112/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_112/agentview/0017.png", "metadata/rollout_112/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_112/agentview/0018.png", "metadata/rollout_112/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_112/agentview/0019.png", "metadata/rollout_112/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_112/agentview/0020.png", "metadata/rollout_112/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_112/agentview/0021.png", "metadata/rollout_112/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_112/agentview/0022.png", "metadata/rollout_112/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_112/agentview/0023.png", "metadata/rollout_112/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_112/agentview/0024.png", "metadata/rollout_112/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_112/agentview/0025.png", "metadata/rollout_112/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0026.png", "metadata/rollout_112/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0027.png", "metadata/rollout_112/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0028.png", "metadata/rollout_112/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0029.png", "metadata/rollout_112/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0030.png", "metadata/rollout_112/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0031.png", "metadata/rollout_112/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_112/agentview/0032.png", "metadata/rollout_112/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0033.png", "metadata/rollout_112/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0034.png", "metadata/rollout_112/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0035.png", "metadata/rollout_112/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0036.png", "metadata/rollout_112/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0037.png", "metadata/rollout_112/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0038.png", "metadata/rollout_112/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0039.png", "metadata/rollout_112/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0040.png", "metadata/rollout_112/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_112/agentview/0041.png", "metadata/rollout_112/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_112/agentview/0042.png", "metadata/rollout_112/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_112/agentview/0043.png", "metadata/rollout_112/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0044.png", "metadata/rollout_112/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0045.png", "metadata/rollout_112/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0046.png", "metadata/rollout_112/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0047.png", "metadata/rollout_112/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0048.png", "metadata/rollout_112/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0049.png", "metadata/rollout_112/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0050.png", "metadata/rollout_112/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0051.png", "metadata/rollout_112/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0052.png", "metadata/rollout_112/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0053.png", "metadata/rollout_112/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0054.png", "metadata/rollout_112/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0055.png", "metadata/rollout_112/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_112/agentview/0056.png", "metadata/rollout_112/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0057.png", "metadata/rollout_112/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0058.png", "metadata/rollout_112/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0059.png", "metadata/rollout_112/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_112/agentview/0060.png", "metadata/rollout_112/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_112/agentview/0061.png", "metadata/rollout_112/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0062.png", "metadata/rollout_112/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_112/agentview/0063.png", "metadata/rollout_112/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_112/agentview/0063.png", "metadata/rollout_112/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0000.png", "metadata/rollout_113/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0001.png", "metadata/rollout_113/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0002.png", "metadata/rollout_113/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0003.png", "metadata/rollout_113/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0004.png", "metadata/rollout_113/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0005.png", "metadata/rollout_113/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0006.png", "metadata/rollout_113/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0007.png", "metadata/rollout_113/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0008.png", "metadata/rollout_113/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0009.png", "metadata/rollout_113/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0010.png", "metadata/rollout_113/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0011.png", "metadata/rollout_113/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0012.png", "metadata/rollout_113/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0013.png", "metadata/rollout_113/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0014.png", "metadata/rollout_113/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0015.png", "metadata/rollout_113/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0016.png", "metadata/rollout_113/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0017.png", "metadata/rollout_113/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_113/agentview/0018.png", "metadata/rollout_113/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_113/agentview/0019.png", "metadata/rollout_113/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_113/agentview/0020.png", "metadata/rollout_113/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_113/agentview/0021.png", "metadata/rollout_113/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_113/agentview/0022.png", "metadata/rollout_113/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_113/agentview/0023.png", "metadata/rollout_113/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_113/agentview/0024.png", "metadata/rollout_113/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0025.png", "metadata/rollout_113/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0026.png", "metadata/rollout_113/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0027.png", "metadata/rollout_113/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0028.png", "metadata/rollout_113/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0029.png", "metadata/rollout_113/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0030.png", "metadata/rollout_113/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_113/agentview/0031.png", "metadata/rollout_113/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0032.png", "metadata/rollout_113/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0033.png", "metadata/rollout_113/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0034.png", "metadata/rollout_113/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0035.png", "metadata/rollout_113/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0036.png", "metadata/rollout_113/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0037.png", "metadata/rollout_113/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0038.png", "metadata/rollout_113/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0039.png", "metadata/rollout_113/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0040.png", "metadata/rollout_113/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0041.png", "metadata/rollout_113/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0042.png", "metadata/rollout_113/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0043.png", "metadata/rollout_113/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0044.png", "metadata/rollout_113/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0045.png", "metadata/rollout_113/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0046.png", "metadata/rollout_113/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0047.png", "metadata/rollout_113/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0048.png", "metadata/rollout_113/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0049.png", "metadata/rollout_113/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0050.png", "metadata/rollout_113/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0051.png", "metadata/rollout_113/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0052.png", "metadata/rollout_113/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_113/agentview/0053.png", "metadata/rollout_113/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_113/agentview/0054.png", "metadata/rollout_113/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0055.png", "metadata/rollout_113/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0056.png", "metadata/rollout_113/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0057.png", "metadata/rollout_113/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_113/agentview/0058.png", "metadata/rollout_113/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_113/agentview/0059.png", "metadata/rollout_113/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0060.png", "metadata/rollout_113/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_113/agentview/0061.png", "metadata/rollout_113/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_113/agentview/0061.png", "metadata/rollout_113/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0000.png", "metadata/rollout_114/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0001.png", "metadata/rollout_114/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0002.png", "metadata/rollout_114/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0003.png", "metadata/rollout_114/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0004.png", "metadata/rollout_114/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0005.png", "metadata/rollout_114/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0006.png", "metadata/rollout_114/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0007.png", "metadata/rollout_114/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0008.png", "metadata/rollout_114/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_114/agentview/0009.png", "metadata/rollout_114/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_114/agentview/0010.png", "metadata/rollout_114/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_114/agentview/0011.png", "metadata/rollout_114/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_114/agentview/0012.png", "metadata/rollout_114/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_114/agentview/0013.png", "metadata/rollout_114/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_114/agentview/0014.png", "metadata/rollout_114/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_114/agentview/0015.png", "metadata/rollout_114/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_114/agentview/0016.png", "metadata/rollout_114/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_114/agentview/0017.png", "metadata/rollout_114/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_114/agentview/0018.png", "metadata/rollout_114/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_114/agentview/0019.png", "metadata/rollout_114/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_114/agentview/0020.png", "metadata/rollout_114/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0021.png", "metadata/rollout_114/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0022.png", "metadata/rollout_114/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0023.png", "metadata/rollout_114/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0024.png", "metadata/rollout_114/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0025.png", "metadata/rollout_114/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0026.png", "metadata/rollout_114/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_114/agentview/0027.png", "metadata/rollout_114/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0028.png", "metadata/rollout_114/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0029.png", "metadata/rollout_114/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0030.png", "metadata/rollout_114/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0031.png", "metadata/rollout_114/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0032.png", "metadata/rollout_114/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0033.png", "metadata/rollout_114/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0034.png", "metadata/rollout_114/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0035.png", "metadata/rollout_114/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_114/agentview/0036.png", "metadata/rollout_114/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_114/agentview/0037.png", "metadata/rollout_114/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_114/agentview/0038.png", "metadata/rollout_114/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_114/agentview/0039.png", "metadata/rollout_114/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_114/agentview/0040.png", "metadata/rollout_114/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_114/agentview/0041.png", "metadata/rollout_114/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_114/agentview/0042.png", "metadata/rollout_114/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_114/agentview/0043.png", "metadata/rollout_114/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_114/agentview/0044.png", "metadata/rollout_114/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_114/agentview/0045.png", "metadata/rollout_114/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0046.png", "metadata/rollout_114/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0047.png", "metadata/rollout_114/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0048.png", "metadata/rollout_114/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_114/agentview/0049.png", "metadata/rollout_114/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_114/agentview/0050.png", "metadata/rollout_114/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0051.png", "metadata/rollout_114/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_114/agentview/0052.png", "metadata/rollout_114/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_114/agentview/0052.png", "metadata/rollout_114/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0000.png", "metadata/rollout_115/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0001.png", "metadata/rollout_115/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0002.png", "metadata/rollout_115/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0003.png", "metadata/rollout_115/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0004.png", "metadata/rollout_115/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0005.png", "metadata/rollout_115/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0006.png", "metadata/rollout_115/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0007.png", "metadata/rollout_115/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0008.png", "metadata/rollout_115/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_115/agentview/0009.png", "metadata/rollout_115/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_115/agentview/0010.png", "metadata/rollout_115/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_115/agentview/0011.png", "metadata/rollout_115/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_115/agentview/0012.png", "metadata/rollout_115/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_115/agentview/0013.png", "metadata/rollout_115/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0014.png", "metadata/rollout_115/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0015.png", "metadata/rollout_115/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0016.png", "metadata/rollout_115/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0017.png", "metadata/rollout_115/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0018.png", "metadata/rollout_115/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0019.png", "metadata/rollout_115/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_115/agentview/0020.png", "metadata/rollout_115/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0021.png", "metadata/rollout_115/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0022.png", "metadata/rollout_115/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0023.png", "metadata/rollout_115/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0024.png", "metadata/rollout_115/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0025.png", "metadata/rollout_115/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0026.png", "metadata/rollout_115/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0027.png", "metadata/rollout_115/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0028.png", "metadata/rollout_115/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0029.png", "metadata/rollout_115/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0030.png", "metadata/rollout_115/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0031.png", "metadata/rollout_115/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0032.png", "metadata/rollout_115/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0033.png", "metadata/rollout_115/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0034.png", "metadata/rollout_115/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0035.png", "metadata/rollout_115/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0036.png", "metadata/rollout_115/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0037.png", "metadata/rollout_115/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0038.png", "metadata/rollout_115/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_115/agentview/0039.png", "metadata/rollout_115/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0040.png", "metadata/rollout_115/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0041.png", "metadata/rollout_115/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0042.png", "metadata/rollout_115/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_115/agentview/0043.png", "metadata/rollout_115/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_115/agentview/0044.png", "metadata/rollout_115/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0045.png", "metadata/rollout_115/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_115/agentview/0046.png", "metadata/rollout_115/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_115/agentview/0046.png", "metadata/rollout_115/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0000.png", "metadata/rollout_116/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0001.png", "metadata/rollout_116/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0002.png", "metadata/rollout_116/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0003.png", "metadata/rollout_116/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0004.png", "metadata/rollout_116/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0005.png", "metadata/rollout_116/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0006.png", "metadata/rollout_116/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0007.png", "metadata/rollout_116/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0008.png", "metadata/rollout_116/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0009.png", "metadata/rollout_116/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_116/agentview/0010.png", "metadata/rollout_116/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0011.png", "metadata/rollout_116/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0012.png", "metadata/rollout_116/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0013.png", "metadata/rollout_116/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0014.png", "metadata/rollout_116/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0015.png", "metadata/rollout_116/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0016.png", "metadata/rollout_116/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0017.png", "metadata/rollout_116/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0018.png", "metadata/rollout_116/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0019.png", "metadata/rollout_116/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_116/agentview/0020.png", "metadata/rollout_116/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_116/agentview/0021.png", "metadata/rollout_116/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_116/agentview/0022.png", "metadata/rollout_116/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_116/agentview/0023.png", "metadata/rollout_116/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_116/agentview/0024.png", "metadata/rollout_116/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_116/agentview/0025.png", "metadata/rollout_116/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0026.png", "metadata/rollout_116/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0027.png", "metadata/rollout_116/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0028.png", "metadata/rollout_116/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0029.png", "metadata/rollout_116/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0030.png", "metadata/rollout_116/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_116/agentview/0031.png", "metadata/rollout_116/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_116/agentview/0032.png", "metadata/rollout_116/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_116/agentview/0033.png", "metadata/rollout_116/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_116/agentview/0034.png", "metadata/rollout_116/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_116/agentview/0035.png", "metadata/rollout_116/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_116/agentview/0036.png", "metadata/rollout_116/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_116/agentview/0037.png", "metadata/rollout_116/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_116/agentview/0038.png", "metadata/rollout_116/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0039.png", "metadata/rollout_116/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0040.png", "metadata/rollout_116/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0041.png", "metadata/rollout_116/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0042.png", "metadata/rollout_116/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0043.png", "metadata/rollout_116/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0044.png", "metadata/rollout_116/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0045.png", "metadata/rollout_116/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0046.png", "metadata/rollout_116/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0047.png", "metadata/rollout_116/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0048.png", "metadata/rollout_116/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0049.png", "metadata/rollout_116/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0050.png", "metadata/rollout_116/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0051.png", "metadata/rollout_116/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0052.png", "metadata/rollout_116/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_116/agentview/0053.png", "metadata/rollout_116/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_116/agentview/0054.png", "metadata/rollout_116/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_116/agentview/0055.png", "metadata/rollout_116/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_116/agentview/0056.png", "metadata/rollout_116/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_116/agentview/0057.png", "metadata/rollout_116/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_116/agentview/0058.png", "metadata/rollout_116/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0059.png", "metadata/rollout_116/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0060.png", "metadata/rollout_116/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_116/agentview/0061.png", "metadata/rollout_116/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_116/agentview/0062.png", "metadata/rollout_116/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_116/agentview/0063.png", "metadata/rollout_116/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_116/agentview/0064.png", "metadata/rollout_116/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_116/agentview/0064.png", "metadata/rollout_116/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0000.png", "metadata/rollout_117/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0001.png", "metadata/rollout_117/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0002.png", "metadata/rollout_117/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0003.png", "metadata/rollout_117/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0004.png", "metadata/rollout_117/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0005.png", "metadata/rollout_117/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0006.png", "metadata/rollout_117/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0007.png", "metadata/rollout_117/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0008.png", "metadata/rollout_117/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_117/agentview/0009.png", "metadata/rollout_117/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0010.png", "metadata/rollout_117/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0011.png", "metadata/rollout_117/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0012.png", "metadata/rollout_117/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0013.png", "metadata/rollout_117/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0014.png", "metadata/rollout_117/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0015.png", "metadata/rollout_117/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0016.png", "metadata/rollout_117/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0017.png", "metadata/rollout_117/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0018.png", "metadata/rollout_117/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_117/agentview/0019.png", "metadata/rollout_117/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_117/agentview/0020.png", "metadata/rollout_117/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_117/agentview/0021.png", "metadata/rollout_117/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_117/agentview/0022.png", "metadata/rollout_117/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_117/agentview/0023.png", "metadata/rollout_117/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_117/agentview/0024.png", "metadata/rollout_117/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_117/agentview/0025.png", "metadata/rollout_117/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_117/agentview/0026.png", "metadata/rollout_117/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_117/agentview/0027.png", "metadata/rollout_117/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0028.png", "metadata/rollout_117/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0029.png", "metadata/rollout_117/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0030.png", "metadata/rollout_117/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0031.png", "metadata/rollout_117/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0032.png", "metadata/rollout_117/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0033.png", "metadata/rollout_117/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_117/agentview/0034.png", "metadata/rollout_117/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0035.png", "metadata/rollout_117/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0036.png", "metadata/rollout_117/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0037.png", "metadata/rollout_117/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0038.png", "metadata/rollout_117/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0039.png", "metadata/rollout_117/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0040.png", "metadata/rollout_117/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0041.png", "metadata/rollout_117/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0042.png", "metadata/rollout_117/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0043.png", "metadata/rollout_117/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0044.png", "metadata/rollout_117/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0045.png", "metadata/rollout_117/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0046.png", "metadata/rollout_117/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0047.png", "metadata/rollout_117/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0048.png", "metadata/rollout_117/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0049.png", "metadata/rollout_117/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0050.png", "metadata/rollout_117/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0051.png", "metadata/rollout_117/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0052.png", "metadata/rollout_117/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0053.png", "metadata/rollout_117/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0054.png", "metadata/rollout_117/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0055.png", "metadata/rollout_117/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0056.png", "metadata/rollout_117/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0057.png", "metadata/rollout_117/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_117/agentview/0058.png", "metadata/rollout_117/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0059.png", "metadata/rollout_117/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0060.png", "metadata/rollout_117/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0061.png", "metadata/rollout_117/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_117/agentview/0062.png", "metadata/rollout_117/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_117/agentview/0063.png", "metadata/rollout_117/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0064.png", "metadata/rollout_117/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_117/agentview/0065.png", "metadata/rollout_117/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_117/agentview/0065.png", "metadata/rollout_117/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0000.png", "metadata/rollout_118/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0001.png", "metadata/rollout_118/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0002.png", "metadata/rollout_118/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0003.png", "metadata/rollout_118/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0004.png", "metadata/rollout_118/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0005.png", "metadata/rollout_118/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0006.png", "metadata/rollout_118/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0007.png", "metadata/rollout_118/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0008.png", "metadata/rollout_118/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_118/agentview/0009.png", "metadata/rollout_118/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0010.png", "metadata/rollout_118/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0011.png", "metadata/rollout_118/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0012.png", "metadata/rollout_118/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0013.png", "metadata/rollout_118/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0014.png", "metadata/rollout_118/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0015.png", "metadata/rollout_118/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0016.png", "metadata/rollout_118/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0017.png", "metadata/rollout_118/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0018.png", "metadata/rollout_118/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_118/agentview/0019.png", "metadata/rollout_118/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_118/agentview/0020.png", "metadata/rollout_118/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_118/agentview/0021.png", "metadata/rollout_118/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0022.png", "metadata/rollout_118/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0023.png", "metadata/rollout_118/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0024.png", "metadata/rollout_118/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0025.png", "metadata/rollout_118/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0026.png", "metadata/rollout_118/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0027.png", "metadata/rollout_118/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_118/agentview/0028.png", "metadata/rollout_118/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0029.png", "metadata/rollout_118/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0030.png", "metadata/rollout_118/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0031.png", "metadata/rollout_118/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0032.png", "metadata/rollout_118/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0033.png", "metadata/rollout_118/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0034.png", "metadata/rollout_118/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0035.png", "metadata/rollout_118/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0036.png", "metadata/rollout_118/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0037.png", "metadata/rollout_118/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0038.png", "metadata/rollout_118/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0039.png", "metadata/rollout_118/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0040.png", "metadata/rollout_118/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0041.png", "metadata/rollout_118/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0042.png", "metadata/rollout_118/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0043.png", "metadata/rollout_118/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0044.png", "metadata/rollout_118/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0045.png", "metadata/rollout_118/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0046.png", "metadata/rollout_118/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_118/agentview/0047.png", "metadata/rollout_118/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0048.png", "metadata/rollout_118/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0049.png", "metadata/rollout_118/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0050.png", "metadata/rollout_118/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_118/agentview/0051.png", "metadata/rollout_118/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_118/agentview/0052.png", "metadata/rollout_118/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0053.png", "metadata/rollout_118/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_118/agentview/0054.png", "metadata/rollout_118/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_118/agentview/0054.png", "metadata/rollout_118/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0000.png", "metadata/rollout_119/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0001.png", "metadata/rollout_119/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0002.png", "metadata/rollout_119/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0003.png", "metadata/rollout_119/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0004.png", "metadata/rollout_119/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0005.png", "metadata/rollout_119/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0006.png", "metadata/rollout_119/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0007.png", "metadata/rollout_119/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0008.png", "metadata/rollout_119/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_119/agentview/0009.png", "metadata/rollout_119/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0010.png", "metadata/rollout_119/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0011.png", "metadata/rollout_119/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0012.png", "metadata/rollout_119/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0013.png", "metadata/rollout_119/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0014.png", "metadata/rollout_119/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0015.png", "metadata/rollout_119/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0016.png", "metadata/rollout_119/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0017.png", "metadata/rollout_119/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0018.png", "metadata/rollout_119/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_119/agentview/0019.png", "metadata/rollout_119/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_119/agentview/0020.png", "metadata/rollout_119/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_119/agentview/0021.png", "metadata/rollout_119/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_119/agentview/0022.png", "metadata/rollout_119/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_119/agentview/0023.png", "metadata/rollout_119/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_119/agentview/0024.png", "metadata/rollout_119/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0025.png", "metadata/rollout_119/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0026.png", "metadata/rollout_119/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0027.png", "metadata/rollout_119/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0028.png", "metadata/rollout_119/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0029.png", "metadata/rollout_119/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0030.png", "metadata/rollout_119/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_119/agentview/0031.png", "metadata/rollout_119/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0032.png", "metadata/rollout_119/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0033.png", "metadata/rollout_119/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0034.png", "metadata/rollout_119/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0035.png", "metadata/rollout_119/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0036.png", "metadata/rollout_119/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0037.png", "metadata/rollout_119/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0038.png", "metadata/rollout_119/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0039.png", "metadata/rollout_119/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0040.png", "metadata/rollout_119/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0041.png", "metadata/rollout_119/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0042.png", "metadata/rollout_119/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0043.png", "metadata/rollout_119/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0044.png", "metadata/rollout_119/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0045.png", "metadata/rollout_119/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0046.png", "metadata/rollout_119/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0047.png", "metadata/rollout_119/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0048.png", "metadata/rollout_119/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0049.png", "metadata/rollout_119/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_119/agentview/0050.png", "metadata/rollout_119/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_119/agentview/0051.png", "metadata/rollout_119/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0052.png", "metadata/rollout_119/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0053.png", "metadata/rollout_119/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0054.png", "metadata/rollout_119/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_119/agentview/0055.png", "metadata/rollout_119/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_119/agentview/0056.png", "metadata/rollout_119/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0057.png", "metadata/rollout_119/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_119/agentview/0058.png", "metadata/rollout_119/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the yellow block in the bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_119/agentview/0058.png", "metadata/rollout_119/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0000.png", "metadata/rollout_120/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0001.png", "metadata/rollout_120/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0002.png", "metadata/rollout_120/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0003.png", "metadata/rollout_120/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0004.png", "metadata/rollout_120/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0005.png", "metadata/rollout_120/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0006.png", "metadata/rollout_120/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0007.png", "metadata/rollout_120/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0008.png", "metadata/rollout_120/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_120/agentview/0009.png", "metadata/rollout_120/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_120/agentview/0010.png", "metadata/rollout_120/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_120/agentview/0011.png", "metadata/rollout_120/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_120/agentview/0012.png", "metadata/rollout_120/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_120/agentview/0013.png", "metadata/rollout_120/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_120/agentview/0014.png", "metadata/rollout_120/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0015.png", "metadata/rollout_120/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0016.png", "metadata/rollout_120/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0017.png", "metadata/rollout_120/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0018.png", "metadata/rollout_120/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0019.png", "metadata/rollout_120/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0020.png", "metadata/rollout_120/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0021.png", "metadata/rollout_120/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0022.png", "metadata/rollout_120/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0023.png", "metadata/rollout_120/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0024.png", "metadata/rollout_120/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0025.png", "metadata/rollout_120/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0026.png", "metadata/rollout_120/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_120/agentview/0027.png", "metadata/rollout_120/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0028.png", "metadata/rollout_120/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0029.png", "metadata/rollout_120/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0030.png", "metadata/rollout_120/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0031.png", "metadata/rollout_120/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0032.png", "metadata/rollout_120/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0033.png", "metadata/rollout_120/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0034.png", "metadata/rollout_120/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0035.png", "metadata/rollout_120/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_120/agentview/0036.png", "metadata/rollout_120/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0037.png", "metadata/rollout_120/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0038.png", "metadata/rollout_120/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0039.png", "metadata/rollout_120/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0040.png", "metadata/rollout_120/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0041.png", "metadata/rollout_120/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0042.png", "metadata/rollout_120/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0043.png", "metadata/rollout_120/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0044.png", "metadata/rollout_120/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0045.png", "metadata/rollout_120/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0046.png", "metadata/rollout_120/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0047.png", "metadata/rollout_120/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0048.png", "metadata/rollout_120/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_120/agentview/0049.png", "metadata/rollout_120/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0050.png", "metadata/rollout_120/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0051.png", "metadata/rollout_120/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_120/agentview/0052.png", "metadata/rollout_120/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0053.png", "metadata/rollout_120/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_120/agentview/0054.png", "metadata/rollout_120/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0055.png", "metadata/rollout_120/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_120/agentview/0056.png", "metadata/rollout_120/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_120/agentview/0056.png", "metadata/rollout_120/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0000.png", "metadata/rollout_121/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0001.png", "metadata/rollout_121/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0002.png", "metadata/rollout_121/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0003.png", "metadata/rollout_121/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0004.png", "metadata/rollout_121/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0005.png", "metadata/rollout_121/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0006.png", "metadata/rollout_121/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0007.png", "metadata/rollout_121/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0008.png", "metadata/rollout_121/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0009.png", "metadata/rollout_121/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0010.png", "metadata/rollout_121/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0011.png", "metadata/rollout_121/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0012.png", "metadata/rollout_121/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0013.png", "metadata/rollout_121/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0014.png", "metadata/rollout_121/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0015.png", "metadata/rollout_121/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0016.png", "metadata/rollout_121/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0017.png", "metadata/rollout_121/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0018.png", "metadata/rollout_121/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0019.png", "metadata/rollout_121/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0020.png", "metadata/rollout_121/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0021.png", "metadata/rollout_121/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0022.png", "metadata/rollout_121/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_121/agentview/0023.png", "metadata/rollout_121/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_121/agentview/0024.png", "metadata/rollout_121/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_121/agentview/0025.png", "metadata/rollout_121/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_121/agentview/0026.png", "metadata/rollout_121/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_121/agentview/0027.png", "metadata/rollout_121/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_121/agentview/0028.png", "metadata/rollout_121/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_121/agentview/0029.png", "metadata/rollout_121/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_121/agentview/0030.png", "metadata/rollout_121/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_121/agentview/0031.png", "metadata/rollout_121/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_121/agentview/0032.png", "metadata/rollout_121/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_121/agentview/0033.png", "metadata/rollout_121/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0034.png", "metadata/rollout_121/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0035.png", "metadata/rollout_121/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0036.png", "metadata/rollout_121/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0037.png", "metadata/rollout_121/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0038.png", "metadata/rollout_121/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0039.png", "metadata/rollout_121/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_121/agentview/0040.png", "metadata/rollout_121/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0041.png", "metadata/rollout_121/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_121/agentview/0042.png", "metadata/rollout_121/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_121/agentview/0043.png", "metadata/rollout_121/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_121/agentview/0044.png", "metadata/rollout_121/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_121/agentview/0045.png", "metadata/rollout_121/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_121/agentview/0045.png", "metadata/rollout_121/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0000.png", "metadata/rollout_122/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0001.png", "metadata/rollout_122/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0002.png", "metadata/rollout_122/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0003.png", "metadata/rollout_122/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0004.png", "metadata/rollout_122/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0005.png", "metadata/rollout_122/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0006.png", "metadata/rollout_122/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0007.png", "metadata/rollout_122/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0008.png", "metadata/rollout_122/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0009.png", "metadata/rollout_122/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0010.png", "metadata/rollout_122/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0011.png", "metadata/rollout_122/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0012.png", "metadata/rollout_122/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0013.png", "metadata/rollout_122/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0014.png", "metadata/rollout_122/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0015.png", "metadata/rollout_122/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0016.png", "metadata/rollout_122/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0017.png", "metadata/rollout_122/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0018.png", "metadata/rollout_122/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0019.png", "metadata/rollout_122/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0020.png", "metadata/rollout_122/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0021.png", "metadata/rollout_122/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_122/agentview/0022.png", "metadata/rollout_122/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_122/agentview/0023.png", "metadata/rollout_122/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_122/agentview/0024.png", "metadata/rollout_122/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_122/agentview/0025.png", "metadata/rollout_122/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0026.png", "metadata/rollout_122/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0027.png", "metadata/rollout_122/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0028.png", "metadata/rollout_122/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0029.png", "metadata/rollout_122/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0030.png", "metadata/rollout_122/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_122/agentview/0031.png", "metadata/rollout_122/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0032.png", "metadata/rollout_122/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0033.png", "metadata/rollout_122/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0034.png", "metadata/rollout_122/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0035.png", "metadata/rollout_122/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0036.png", "metadata/rollout_122/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0037.png", "metadata/rollout_122/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0038.png", "metadata/rollout_122/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0039.png", "metadata/rollout_122/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_122/agentview/0040.png", "metadata/rollout_122/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_122/agentview/0041.png", "metadata/rollout_122/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_122/agentview/0042.png", "metadata/rollout_122/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0043.png", "metadata/rollout_122/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0044.png", "metadata/rollout_122/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0045.png", "metadata/rollout_122/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0046.png", "metadata/rollout_122/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0047.png", "metadata/rollout_122/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0048.png", "metadata/rollout_122/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0049.png", "metadata/rollout_122/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_122/agentview/0050.png", "metadata/rollout_122/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_122/agentview/0051.png", "metadata/rollout_122/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0052.png", "metadata/rollout_122/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0053.png", "metadata/rollout_122/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_122/agentview/0054.png", "metadata/rollout_122/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_122/agentview/0055.png", "metadata/rollout_122/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0056.png", "metadata/rollout_122/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_122/agentview/0057.png", "metadata/rollout_122/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_122/agentview/0057.png", "metadata/rollout_122/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0000.png", "metadata/rollout_123/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0001.png", "metadata/rollout_123/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0002.png", "metadata/rollout_123/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0003.png", "metadata/rollout_123/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0004.png", "metadata/rollout_123/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0005.png", "metadata/rollout_123/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0006.png", "metadata/rollout_123/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0007.png", "metadata/rollout_123/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0008.png", "metadata/rollout_123/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0009.png", "metadata/rollout_123/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0010.png", "metadata/rollout_123/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0011.png", "metadata/rollout_123/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0012.png", "metadata/rollout_123/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0013.png", "metadata/rollout_123/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0014.png", "metadata/rollout_123/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0015.png", "metadata/rollout_123/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0016.png", "metadata/rollout_123/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0017.png", "metadata/rollout_123/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0018.png", "metadata/rollout_123/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0019.png", "metadata/rollout_123/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_123/agentview/0020.png", "metadata/rollout_123/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_123/agentview/0021.png", "metadata/rollout_123/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_123/agentview/0022.png", "metadata/rollout_123/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_123/agentview/0023.png", "metadata/rollout_123/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_123/agentview/0024.png", "metadata/rollout_123/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0025.png", "metadata/rollout_123/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0026.png", "metadata/rollout_123/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0027.png", "metadata/rollout_123/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0028.png", "metadata/rollout_123/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0029.png", "metadata/rollout_123/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_123/agentview/0030.png", "metadata/rollout_123/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0031.png", "metadata/rollout_123/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0032.png", "metadata/rollout_123/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0033.png", "metadata/rollout_123/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0034.png", "metadata/rollout_123/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0035.png", "metadata/rollout_123/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0036.png", "metadata/rollout_123/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0037.png", "metadata/rollout_123/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0038.png", "metadata/rollout_123/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_123/agentview/0039.png", "metadata/rollout_123/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_123/agentview/0040.png", "metadata/rollout_123/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_123/agentview/0041.png", "metadata/rollout_123/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_123/agentview/0042.png", "metadata/rollout_123/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_123/agentview/0043.png", "metadata/rollout_123/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0044.png", "metadata/rollout_123/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0045.png", "metadata/rollout_123/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0046.png", "metadata/rollout_123/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0047.png", "metadata/rollout_123/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0048.png", "metadata/rollout_123/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0049.png", "metadata/rollout_123/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0050.png", "metadata/rollout_123/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_123/agentview/0051.png", "metadata/rollout_123/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0052.png", "metadata/rollout_123/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0053.png", "metadata/rollout_123/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_123/agentview/0054.png", "metadata/rollout_123/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_123/agentview/0055.png", "metadata/rollout_123/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0056.png", "metadata/rollout_123/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_123/agentview/0057.png", "metadata/rollout_123/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_123/agentview/0057.png", "metadata/rollout_123/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0000.png", "metadata/rollout_124/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0001.png", "metadata/rollout_124/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0002.png", "metadata/rollout_124/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0003.png", "metadata/rollout_124/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0004.png", "metadata/rollout_124/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0005.png", "metadata/rollout_124/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0006.png", "metadata/rollout_124/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0007.png", "metadata/rollout_124/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0008.png", "metadata/rollout_124/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0009.png", "metadata/rollout_124/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0010.png", "metadata/rollout_124/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0011.png", "metadata/rollout_124/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0012.png", "metadata/rollout_124/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0013.png", "metadata/rollout_124/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0014.png", "metadata/rollout_124/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0015.png", "metadata/rollout_124/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_124/agentview/0016.png", "metadata/rollout_124/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_124/agentview/0017.png", "metadata/rollout_124/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_124/agentview/0018.png", "metadata/rollout_124/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_124/agentview/0019.png", "metadata/rollout_124/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_124/agentview/0020.png", "metadata/rollout_124/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0021.png", "metadata/rollout_124/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0022.png", "metadata/rollout_124/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0023.png", "metadata/rollout_124/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0024.png", "metadata/rollout_124/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0025.png", "metadata/rollout_124/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0026.png", "metadata/rollout_124/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0027.png", "metadata/rollout_124/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0028.png", "metadata/rollout_124/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_124/agentview/0029.png", "metadata/rollout_124/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_124/agentview/0030.png", "metadata/rollout_124/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_124/agentview/0031.png", "metadata/rollout_124/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_124/agentview/0032.png", "metadata/rollout_124/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0033.png", "metadata/rollout_124/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_BACK, MV_BACK, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0034.png", "metadata/rollout_124/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0035.png", "metadata/rollout_124/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0036.png", "metadata/rollout_124/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0037.png", "metadata/rollout_124/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0038.png", "metadata/rollout_124/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0039.png", "metadata/rollout_124/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_124/agentview/0040.png", "metadata/rollout_124/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0041.png", "metadata/rollout_124/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0042.png", "metadata/rollout_124/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0043.png", "metadata/rollout_124/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0044.png", "metadata/rollout_124/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0045.png", "metadata/rollout_124/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0046.png", "metadata/rollout_124/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0047.png", "metadata/rollout_124/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0048.png", "metadata/rollout_124/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0049.png", "metadata/rollout_124/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0050.png", "metadata/rollout_124/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0051.png", "metadata/rollout_124/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0052.png", "metadata/rollout_124/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_124/agentview/0053.png", "metadata/rollout_124/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0054.png", "metadata/rollout_124/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_124/agentview/0055.png", "metadata/rollout_124/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0056.png", "metadata/rollout_124/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_124/agentview/0057.png", "metadata/rollout_124/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0058.png", "metadata/rollout_124/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_124/agentview/0059.png", "metadata/rollout_124/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_124/agentview/0059.png", "metadata/rollout_124/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0000.png", "metadata/rollout_125/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0001.png", "metadata/rollout_125/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0002.png", "metadata/rollout_125/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0003.png", "metadata/rollout_125/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0004.png", "metadata/rollout_125/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0005.png", "metadata/rollout_125/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0006.png", "metadata/rollout_125/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0007.png", "metadata/rollout_125/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0008.png", "metadata/rollout_125/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0009.png", "metadata/rollout_125/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0010.png", "metadata/rollout_125/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0011.png", "metadata/rollout_125/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0012.png", "metadata/rollout_125/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0013.png", "metadata/rollout_125/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0014.png", "metadata/rollout_125/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0015.png", "metadata/rollout_125/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0016.png", "metadata/rollout_125/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_125/agentview/0017.png", "metadata/rollout_125/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_125/agentview/0018.png", "metadata/rollout_125/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_125/agentview/0019.png", "metadata/rollout_125/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_125/agentview/0020.png", "metadata/rollout_125/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0021.png", "metadata/rollout_125/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0022.png", "metadata/rollout_125/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0023.png", "metadata/rollout_125/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0024.png", "metadata/rollout_125/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0025.png", "metadata/rollout_125/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_125/agentview/0026.png", "metadata/rollout_125/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_125/agentview/0027.png", "metadata/rollout_125/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_125/agentview/0028.png", "metadata/rollout_125/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_125/agentview/0029.png", "metadata/rollout_125/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_125/agentview/0030.png", "metadata/rollout_125/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_125/agentview/0031.png", "metadata/rollout_125/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_125/agentview/0032.png", "metadata/rollout_125/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_125/agentview/0033.png", "metadata/rollout_125/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0034.png", "metadata/rollout_125/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0035.png", "metadata/rollout_125/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0036.png", "metadata/rollout_125/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0037.png", "metadata/rollout_125/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0038.png", "metadata/rollout_125/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0039.png", "metadata/rollout_125/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_125/agentview/0040.png", "metadata/rollout_125/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_125/agentview/0041.png", "metadata/rollout_125/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_125/agentview/0042.png", "metadata/rollout_125/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_125/agentview/0043.png", "metadata/rollout_125/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_125/agentview/0044.png", "metadata/rollout_125/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0045.png", "metadata/rollout_125/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_125/agentview/0046.png", "metadata/rollout_125/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_125/agentview/0047.png", "metadata/rollout_125/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_125/agentview/0048.png", "metadata/rollout_125/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_125/agentview/0049.png", "metadata/rollout_125/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_125/agentview/0049.png", "metadata/rollout_125/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0000.png", "metadata/rollout_126/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0001.png", "metadata/rollout_126/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0002.png", "metadata/rollout_126/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0003.png", "metadata/rollout_126/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0004.png", "metadata/rollout_126/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0005.png", "metadata/rollout_126/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0006.png", "metadata/rollout_126/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0007.png", "metadata/rollout_126/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0008.png", "metadata/rollout_126/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0009.png", "metadata/rollout_126/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0010.png", "metadata/rollout_126/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0011.png", "metadata/rollout_126/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0012.png", "metadata/rollout_126/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0013.png", "metadata/rollout_126/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0014.png", "metadata/rollout_126/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0015.png", "metadata/rollout_126/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_126/agentview/0016.png", "metadata/rollout_126/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0017.png", "metadata/rollout_126/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0018.png", "metadata/rollout_126/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0019.png", "metadata/rollout_126/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0020.png", "metadata/rollout_126/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0021.png", "metadata/rollout_126/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0022.png", "metadata/rollout_126/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_126/agentview/0023.png", "metadata/rollout_126/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0024.png", "metadata/rollout_126/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0025.png", "metadata/rollout_126/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0026.png", "metadata/rollout_126/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0027.png", "metadata/rollout_126/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0028.png", "metadata/rollout_126/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0029.png", "metadata/rollout_126/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0030.png", "metadata/rollout_126/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0031.png", "metadata/rollout_126/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0032.png", "metadata/rollout_126/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0033.png", "metadata/rollout_126/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0034.png", "metadata/rollout_126/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0035.png", "metadata/rollout_126/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0036.png", "metadata/rollout_126/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0037.png", "metadata/rollout_126/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0038.png", "metadata/rollout_126/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0039.png", "metadata/rollout_126/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0040.png", "metadata/rollout_126/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_126/agentview/0041.png", "metadata/rollout_126/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_126/agentview/0042.png", "metadata/rollout_126/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_126/agentview/0043.png", "metadata/rollout_126/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_126/agentview/0044.png", "metadata/rollout_126/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_126/agentview/0045.png", "metadata/rollout_126/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0046.png", "metadata/rollout_126/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0047.png", "metadata/rollout_126/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_126/agentview/0048.png", "metadata/rollout_126/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0049.png", "metadata/rollout_126/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_126/agentview/0050.png", "metadata/rollout_126/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0051.png", "metadata/rollout_126/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_126/agentview/0052.png", "metadata/rollout_126/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_126/agentview/0052.png", "metadata/rollout_126/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0000.png", "metadata/rollout_127/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0001.png", "metadata/rollout_127/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0002.png", "metadata/rollout_127/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0003.png", "metadata/rollout_127/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0004.png", "metadata/rollout_127/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0005.png", "metadata/rollout_127/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0006.png", "metadata/rollout_127/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0007.png", "metadata/rollout_127/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0008.png", "metadata/rollout_127/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0009.png", "metadata/rollout_127/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0010.png", "metadata/rollout_127/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0011.png", "metadata/rollout_127/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0012.png", "metadata/rollout_127/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0013.png", "metadata/rollout_127/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0014.png", "metadata/rollout_127/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0015.png", "metadata/rollout_127/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_127/agentview/0016.png", "metadata/rollout_127/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_127/agentview/0017.png", "metadata/rollout_127/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_127/agentview/0018.png", "metadata/rollout_127/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_127/agentview/0019.png", "metadata/rollout_127/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_127/agentview/0020.png", "metadata/rollout_127/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_127/agentview/0021.png", "metadata/rollout_127/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0022.png", "metadata/rollout_127/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0023.png", "metadata/rollout_127/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0024.png", "metadata/rollout_127/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0025.png", "metadata/rollout_127/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0026.png", "metadata/rollout_127/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_127/agentview/0027.png", "metadata/rollout_127/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0028.png", "metadata/rollout_127/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0029.png", "metadata/rollout_127/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0030.png", "metadata/rollout_127/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0031.png", "metadata/rollout_127/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0032.png", "metadata/rollout_127/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0033.png", "metadata/rollout_127/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0034.png", "metadata/rollout_127/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0035.png", "metadata/rollout_127/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_127/agentview/0036.png", "metadata/rollout_127/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0037.png", "metadata/rollout_127/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0038.png", "metadata/rollout_127/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0039.png", "metadata/rollout_127/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0040.png", "metadata/rollout_127/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0041.png", "metadata/rollout_127/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0042.png", "metadata/rollout_127/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0043.png", "metadata/rollout_127/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_127/agentview/0044.png", "metadata/rollout_127/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0045.png", "metadata/rollout_127/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0046.png", "metadata/rollout_127/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_127/agentview/0047.png", "metadata/rollout_127/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_127/agentview/0048.png", "metadata/rollout_127/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0049.png", "metadata/rollout_127/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_127/agentview/0050.png", "metadata/rollout_127/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_127/agentview/0050.png", "metadata/rollout_127/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0000.png", "metadata/rollout_128/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0001.png", "metadata/rollout_128/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0002.png", "metadata/rollout_128/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0003.png", "metadata/rollout_128/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0004.png", "metadata/rollout_128/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0005.png", "metadata/rollout_128/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0006.png", "metadata/rollout_128/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0007.png", "metadata/rollout_128/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0008.png", "metadata/rollout_128/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0009.png", "metadata/rollout_128/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0010.png", "metadata/rollout_128/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0011.png", "metadata/rollout_128/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0012.png", "metadata/rollout_128/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0013.png", "metadata/rollout_128/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0014.png", "metadata/rollout_128/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0015.png", "metadata/rollout_128/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0016.png", "metadata/rollout_128/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0017.png", "metadata/rollout_128/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0018.png", "metadata/rollout_128/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_128/agentview/0019.png", "metadata/rollout_128/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_128/agentview/0020.png", "metadata/rollout_128/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_128/agentview/0021.png", "metadata/rollout_128/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_128/agentview/0022.png", "metadata/rollout_128/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_128/agentview/0023.png", "metadata/rollout_128/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_128/agentview/0024.png", "metadata/rollout_128/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0025.png", "metadata/rollout_128/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0026.png", "metadata/rollout_128/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0027.png", "metadata/rollout_128/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0028.png", "metadata/rollout_128/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0029.png", "metadata/rollout_128/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_128/agentview/0030.png", "metadata/rollout_128/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0031.png", "metadata/rollout_128/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0032.png", "metadata/rollout_128/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0033.png", "metadata/rollout_128/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0034.png", "metadata/rollout_128/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0035.png", "metadata/rollout_128/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0036.png", "metadata/rollout_128/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0037.png", "metadata/rollout_128/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0038.png", "metadata/rollout_128/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0039.png", "metadata/rollout_128/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0040.png", "metadata/rollout_128/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0041.png", "metadata/rollout_128/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0042.png", "metadata/rollout_128/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0043.png", "metadata/rollout_128/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0044.png", "metadata/rollout_128/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0045.png", "metadata/rollout_128/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_128/agentview/0046.png", "metadata/rollout_128/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0047.png", "metadata/rollout_128/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0048.png", "metadata/rollout_128/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_128/agentview/0049.png", "metadata/rollout_128/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_128/agentview/0050.png", "metadata/rollout_128/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0051.png", "metadata/rollout_128/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_128/agentview/0052.png", "metadata/rollout_128/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_128/agentview/0052.png", "metadata/rollout_128/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0000.png", "metadata/rollout_129/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0001.png", "metadata/rollout_129/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0002.png", "metadata/rollout_129/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0003.png", "metadata/rollout_129/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0004.png", "metadata/rollout_129/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0005.png", "metadata/rollout_129/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0006.png", "metadata/rollout_129/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0007.png", "metadata/rollout_129/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0008.png", "metadata/rollout_129/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0009.png", "metadata/rollout_129/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0010.png", "metadata/rollout_129/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0011.png", "metadata/rollout_129/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0012.png", "metadata/rollout_129/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0013.png", "metadata/rollout_129/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0014.png", "metadata/rollout_129/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0015.png", "metadata/rollout_129/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0016.png", "metadata/rollout_129/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0017.png", "metadata/rollout_129/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0018.png", "metadata/rollout_129/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0019.png", "metadata/rollout_129/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0020.png", "metadata/rollout_129/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_129/agentview/0021.png", "metadata/rollout_129/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_129/agentview/0022.png", "metadata/rollout_129/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_129/agentview/0023.png", "metadata/rollout_129/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0024.png", "metadata/rollout_129/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0025.png", "metadata/rollout_129/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0026.png", "metadata/rollout_129/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0027.png", "metadata/rollout_129/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0028.png", "metadata/rollout_129/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_129/agentview/0029.png", "metadata/rollout_129/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0030.png", "metadata/rollout_129/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0031.png", "metadata/rollout_129/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0032.png", "metadata/rollout_129/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0033.png", "metadata/rollout_129/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0034.png", "metadata/rollout_129/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0035.png", "metadata/rollout_129/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0036.png", "metadata/rollout_129/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0037.png", "metadata/rollout_129/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0038.png", "metadata/rollout_129/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0039.png", "metadata/rollout_129/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0040.png", "metadata/rollout_129/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0041.png", "metadata/rollout_129/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0042.png", "metadata/rollout_129/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0043.png", "metadata/rollout_129/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0044.png", "metadata/rollout_129/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0045.png", "metadata/rollout_129/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0046.png", "metadata/rollout_129/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_129/agentview/0047.png", "metadata/rollout_129/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_129/agentview/0048.png", "metadata/rollout_129/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_129/agentview/0049.png", "metadata/rollout_129/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_129/agentview/0050.png", "metadata/rollout_129/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_FWD, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_129/agentview/0051.png", "metadata/rollout_129/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0052.png", "metadata/rollout_129/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0053.png", "metadata/rollout_129/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_129/agentview/0054.png", "metadata/rollout_129/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0055.png", "metadata/rollout_129/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_129/agentview/0056.png", "metadata/rollout_129/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0057.png", "metadata/rollout_129/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_129/agentview/0058.png", "metadata/rollout_129/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: Put the small red yogurt in the red bowl\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_129/agentview/0058.png", "metadata/rollout_129/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_130/agentview/0000.png", "metadata/rollout_130/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_130/agentview/0001.png", "metadata/rollout_130/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_130/agentview/0002.png", "metadata/rollout_130/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_130/agentview/0003.png", "metadata/rollout_130/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_130/agentview/0004.png", "metadata/rollout_130/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_130/agentview/0005.png", "metadata/rollout_130/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_130/agentview/0006.png", "metadata/rollout_130/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_130/agentview/0007.png", "metadata/rollout_130/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_130/agentview/0008.png", "metadata/rollout_130/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_130/agentview/0009.png", "metadata/rollout_130/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_130/agentview/0010.png", "metadata/rollout_130/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_130/agentview/0011.png", "metadata/rollout_130/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_130/agentview/0012.png", "metadata/rollout_130/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_130/agentview/0013.png", "metadata/rollout_130/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_130/agentview/0014.png", "metadata/rollout_130/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0015.png", "metadata/rollout_130/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0016.png", "metadata/rollout_130/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0017.png", "metadata/rollout_130/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0018.png", "metadata/rollout_130/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0019.png", "metadata/rollout_130/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0020.png", "metadata/rollout_130/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0021.png", "metadata/rollout_130/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0022.png", "metadata/rollout_130/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0023.png", "metadata/rollout_130/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_130/agentview/0024.png", "metadata/rollout_130/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0025.png", "metadata/rollout_130/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0026.png", "metadata/rollout_130/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0027.png", "metadata/rollout_130/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_130/agentview/0028.png", "metadata/rollout_130/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0029.png", "metadata/rollout_130/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0030.png", "metadata/rollout_130/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0031.png", "metadata/rollout_130/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0032.png", "metadata/rollout_130/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0033.png", "metadata/rollout_130/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0034.png", "metadata/rollout_130/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_130/agentview/0035.png", "metadata/rollout_130/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_130/agentview/0036.png", "metadata/rollout_130/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0037.png", "metadata/rollout_130/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_130/agentview/0038.png", "metadata/rollout_130/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_LEFT, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_130/agentview/0039.png", "metadata/rollout_130/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_130/agentview/0040.png", "metadata/rollout_130/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_130/agentview/0041.png", "metadata/rollout_130/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_130/agentview/0042.png", "metadata/rollout_130/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_130/agentview/0043.png", "metadata/rollout_130/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_130/agentview/0044.png", "metadata/rollout_130/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0045.png", "metadata/rollout_130/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0046.png", "metadata/rollout_130/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0047.png", "metadata/rollout_130/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0048.png", "metadata/rollout_130/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_130/agentview/0049.png", "metadata/rollout_130/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_130/agentview/0050.png", "metadata/rollout_130/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_130/agentview/0051.png", "metadata/rollout_130/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0052.png", "metadata/rollout_130/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0053.png", "metadata/rollout_130/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0054.png", "metadata/rollout_130/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0055.png", "metadata/rollout_130/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_130/agentview/0056.png", "metadata/rollout_130/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_130/agentview/0057.png", "metadata/rollout_130/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_130/agentview/0057.png", "metadata/rollout_130/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0000.png", "metadata/rollout_131/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0001.png", "metadata/rollout_131/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0002.png", "metadata/rollout_131/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_131/agentview/0003.png", "metadata/rollout_131/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_131/agentview/0004.png", "metadata/rollout_131/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_131/agentview/0005.png", "metadata/rollout_131/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0006.png", "metadata/rollout_131/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0007.png", "metadata/rollout_131/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0008.png", "metadata/rollout_131/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0009.png", "metadata/rollout_131/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0010.png", "metadata/rollout_131/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_131/agentview/0011.png", "metadata/rollout_131/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_131/agentview/0012.png", "metadata/rollout_131/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_131/agentview/0013.png", "metadata/rollout_131/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0014.png", "metadata/rollout_131/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0015.png", "metadata/rollout_131/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0016.png", "metadata/rollout_131/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0017.png", "metadata/rollout_131/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0018.png", "metadata/rollout_131/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0019.png", "metadata/rollout_131/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_131/agentview/0020.png", "metadata/rollout_131/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_131/agentview/0021.png", "metadata/rollout_131/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_131/agentview/0022.png", "metadata/rollout_131/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_131/agentview/0023.png", "metadata/rollout_131/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_131/agentview/0024.png", "metadata/rollout_131/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_131/agentview/0025.png", "metadata/rollout_131/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0026.png", "metadata/rollout_131/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0027.png", "metadata/rollout_131/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_131/agentview/0028.png", "metadata/rollout_131/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0029.png", "metadata/rollout_131/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_131/agentview/0030.png", "metadata/rollout_131/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_131/agentview/0031.png", "metadata/rollout_131/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_131/agentview/0032.png", "metadata/rollout_131/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_131/agentview/0033.png", "metadata/rollout_131/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_131/agentview/0034.png", "metadata/rollout_131/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_131/agentview/0035.png", "metadata/rollout_131/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_131/agentview/0036.png", "metadata/rollout_131/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0037.png", "metadata/rollout_131/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0038.png", "metadata/rollout_131/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0039.png", "metadata/rollout_131/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0040.png", "metadata/rollout_131/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0041.png", "metadata/rollout_131/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0042.png", "metadata/rollout_131/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0043.png", "metadata/rollout_131/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_131/agentview/0044.png", "metadata/rollout_131/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0045.png", "metadata/rollout_131/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0046.png", "metadata/rollout_131/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0047.png", "metadata/rollout_131/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_131/agentview/0048.png", "metadata/rollout_131/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_131/agentview/0049.png", "metadata/rollout_131/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_131/agentview/0050.png", "metadata/rollout_131/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_131/agentview/0051.png", "metadata/rollout_131/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_131/agentview/0052.png", "metadata/rollout_131/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_131/agentview/0053.png", "metadata/rollout_131/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_131/agentview/0053.png", "metadata/rollout_131/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0000.png", "metadata/rollout_132/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0001.png", "metadata/rollout_132/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_132/agentview/0002.png", "metadata/rollout_132/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0003.png", "metadata/rollout_132/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0004.png", "metadata/rollout_132/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0005.png", "metadata/rollout_132/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0006.png", "metadata/rollout_132/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0007.png", "metadata/rollout_132/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0008.png", "metadata/rollout_132/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0009.png", "metadata/rollout_132/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0010.png", "metadata/rollout_132/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0011.png", "metadata/rollout_132/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0012.png", "metadata/rollout_132/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0013.png", "metadata/rollout_132/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0014.png", "metadata/rollout_132/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0015.png", "metadata/rollout_132/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0016.png", "metadata/rollout_132/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0017.png", "metadata/rollout_132/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0018.png", "metadata/rollout_132/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0019.png", "metadata/rollout_132/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_132/agentview/0020.png", "metadata/rollout_132/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0021.png", "metadata/rollout_132/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0022.png", "metadata/rollout_132/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0023.png", "metadata/rollout_132/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0024.png", "metadata/rollout_132/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0025.png", "metadata/rollout_132/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0026.png", "metadata/rollout_132/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0027.png", "metadata/rollout_132/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0028.png", "metadata/rollout_132/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_132/agentview/0029.png", "metadata/rollout_132/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_132/agentview/0030.png", "metadata/rollout_132/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0031.png", "metadata/rollout_132/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0032.png", "metadata/rollout_132/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0033.png", "metadata/rollout_132/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0034.png", "metadata/rollout_132/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0035.png", "metadata/rollout_132/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_132/agentview/0036.png", "metadata/rollout_132/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_132/agentview/0037.png", "metadata/rollout_132/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_132/agentview/0038.png", "metadata/rollout_132/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_132/agentview/0039.png", "metadata/rollout_132/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0040.png", "metadata/rollout_132/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0041.png", "metadata/rollout_132/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0042.png", "metadata/rollout_132/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_132/agentview/0043.png", "metadata/rollout_132/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_132/agentview/0044.png", "metadata/rollout_132/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_132/agentview/0045.png", "metadata/rollout_132/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0046.png", "metadata/rollout_132/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0047.png", "metadata/rollout_132/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0048.png", "metadata/rollout_132/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_132/agentview/0049.png", "metadata/rollout_132/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_132/agentview/0049.png", "metadata/rollout_132/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0000.png", "metadata/rollout_133/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0001.png", "metadata/rollout_133/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0002.png", "metadata/rollout_133/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0003.png", "metadata/rollout_133/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0004.png", "metadata/rollout_133/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_133/agentview/0005.png", "metadata/rollout_133/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_133/agentview/0006.png", "metadata/rollout_133/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0007.png", "metadata/rollout_133/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0008.png", "metadata/rollout_133/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0009.png", "metadata/rollout_133/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0010.png", "metadata/rollout_133/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0011.png", "metadata/rollout_133/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0012.png", "metadata/rollout_133/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_133/agentview/0013.png", "metadata/rollout_133/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0014.png", "metadata/rollout_133/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0015.png", "metadata/rollout_133/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0016.png", "metadata/rollout_133/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0017.png", "metadata/rollout_133/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0018.png", "metadata/rollout_133/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0019.png", "metadata/rollout_133/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_133/agentview/0020.png", "metadata/rollout_133/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_133/agentview/0021.png", "metadata/rollout_133/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_133/agentview/0022.png", "metadata/rollout_133/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_133/agentview/0023.png", "metadata/rollout_133/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_133/agentview/0024.png", "metadata/rollout_133/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_133/agentview/0025.png", "metadata/rollout_133/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0026.png", "metadata/rollout_133/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0027.png", "metadata/rollout_133/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_133/agentview/0028.png", "metadata/rollout_133/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_133/agentview/0029.png", "metadata/rollout_133/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0030.png", "metadata/rollout_133/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0031.png", "metadata/rollout_133/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0032.png", "metadata/rollout_133/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_133/agentview/0033.png", "metadata/rollout_133/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_133/agentview/0034.png", "metadata/rollout_133/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0035.png", "metadata/rollout_133/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_133/agentview/0036.png", "metadata/rollout_133/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0037.png", "metadata/rollout_133/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0038.png", "metadata/rollout_133/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0039.png", "metadata/rollout_133/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_133/agentview/0040.png", "metadata/rollout_133/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_133/agentview/0041.png", "metadata/rollout_133/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_133/agentview/0042.png", "metadata/rollout_133/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_133/agentview/0043.png", "metadata/rollout_133/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_133/agentview/0044.png", "metadata/rollout_133/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_133/agentview/0045.png", "metadata/rollout_133/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_133/agentview/0045.png", "metadata/rollout_133/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0000.png", "metadata/rollout_134/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0001.png", "metadata/rollout_134/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0002.png", "metadata/rollout_134/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0003.png", "metadata/rollout_134/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0004.png", "metadata/rollout_134/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0005.png", "metadata/rollout_134/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_134/agentview/0006.png", "metadata/rollout_134/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_134/agentview/0007.png", "metadata/rollout_134/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_134/agentview/0008.png", "metadata/rollout_134/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0009.png", "metadata/rollout_134/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0010.png", "metadata/rollout_134/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0011.png", "metadata/rollout_134/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0012.png", "metadata/rollout_134/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0013.png", "metadata/rollout_134/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0014.png", "metadata/rollout_134/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0015.png", "metadata/rollout_134/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0016.png", "metadata/rollout_134/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0017.png", "metadata/rollout_134/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0018.png", "metadata/rollout_134/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0019.png", "metadata/rollout_134/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0020.png", "metadata/rollout_134/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0021.png", "metadata/rollout_134/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0022.png", "metadata/rollout_134/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_134/agentview/0023.png", "metadata/rollout_134/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0024.png", "metadata/rollout_134/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0025.png", "metadata/rollout_134/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0026.png", "metadata/rollout_134/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0027.png", "metadata/rollout_134/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_134/agentview/0028.png", "metadata/rollout_134/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0029.png", "metadata/rollout_134/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0030.png", "metadata/rollout_134/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0031.png", "metadata/rollout_134/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0032.png", "metadata/rollout_134/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0033.png", "metadata/rollout_134/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0034.png", "metadata/rollout_134/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_134/agentview/0035.png", "metadata/rollout_134/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_134/agentview/0036.png", "metadata/rollout_134/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0037.png", "metadata/rollout_134/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_134/agentview/0038.png", "metadata/rollout_134/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_134/agentview/0039.png", "metadata/rollout_134/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_134/agentview/0040.png", "metadata/rollout_134/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_134/agentview/0041.png", "metadata/rollout_134/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_134/agentview/0042.png", "metadata/rollout_134/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0043.png", "metadata/rollout_134/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0044.png", "metadata/rollout_134/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0045.png", "metadata/rollout_134/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0046.png", "metadata/rollout_134/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_134/agentview/0047.png", "metadata/rollout_134/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_134/agentview/0048.png", "metadata/rollout_134/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_134/agentview/0049.png", "metadata/rollout_134/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0050.png", "metadata/rollout_134/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0051.png", "metadata/rollout_134/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0052.png", "metadata/rollout_134/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0053.png", "metadata/rollout_134/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_134/agentview/0054.png", "metadata/rollout_134/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_134/agentview/0054.png", "metadata/rollout_134/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0000.png", "metadata/rollout_135/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0001.png", "metadata/rollout_135/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0002.png", "metadata/rollout_135/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0003.png", "metadata/rollout_135/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0004.png", "metadata/rollout_135/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0005.png", "metadata/rollout_135/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0006.png", "metadata/rollout_135/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0007.png", "metadata/rollout_135/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0008.png", "metadata/rollout_135/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0009.png", "metadata/rollout_135/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0010.png", "metadata/rollout_135/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_135/agentview/0011.png", "metadata/rollout_135/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_135/agentview/0012.png", "metadata/rollout_135/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_135/agentview/0013.png", "metadata/rollout_135/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_135/agentview/0014.png", "metadata/rollout_135/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_135/agentview/0015.png", "metadata/rollout_135/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0016.png", "metadata/rollout_135/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0017.png", "metadata/rollout_135/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0018.png", "metadata/rollout_135/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0019.png", "metadata/rollout_135/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0020.png", "metadata/rollout_135/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0021.png", "metadata/rollout_135/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0022.png", "metadata/rollout_135/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_135/agentview/0023.png", "metadata/rollout_135/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_135/agentview/0024.png", "metadata/rollout_135/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_135/agentview/0025.png", "metadata/rollout_135/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_135/agentview/0026.png", "metadata/rollout_135/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_135/agentview/0027.png", "metadata/rollout_135/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_135/agentview/0028.png", "metadata/rollout_135/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0029.png", "metadata/rollout_135/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_135/agentview/0030.png", "metadata/rollout_135/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0031.png", "metadata/rollout_135/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0032.png", "metadata/rollout_135/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0033.png", "metadata/rollout_135/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0034.png", "metadata/rollout_135/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0035.png", "metadata/rollout_135/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_135/agentview/0036.png", "metadata/rollout_135/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_135/agentview/0037.png", "metadata/rollout_135/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_135/agentview/0038.png", "metadata/rollout_135/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_135/agentview/0039.png", "metadata/rollout_135/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_135/agentview/0040.png", "metadata/rollout_135/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_135/agentview/0041.png", "metadata/rollout_135/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_135/agentview/0042.png", "metadata/rollout_135/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_135/agentview/0043.png", "metadata/rollout_135/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0044.png", "metadata/rollout_135/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0045.png", "metadata/rollout_135/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0046.png", "metadata/rollout_135/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_135/agentview/0047.png", "metadata/rollout_135/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_135/agentview/0048.png", "metadata/rollout_135/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_135/agentview/0049.png", "metadata/rollout_135/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_135/agentview/0050.png", "metadata/rollout_135/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_135/agentview/0051.png", "metadata/rollout_135/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_135/agentview/0052.png", "metadata/rollout_135/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_135/agentview/0052.png", "metadata/rollout_135/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_136/agentview/0000.png", "metadata/rollout_136/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_136/agentview/0001.png", "metadata/rollout_136/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_136/agentview/0002.png", "metadata/rollout_136/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_136/agentview/0003.png", "metadata/rollout_136/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0004.png", "metadata/rollout_136/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0005.png", "metadata/rollout_136/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0006.png", "metadata/rollout_136/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0007.png", "metadata/rollout_136/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0008.png", "metadata/rollout_136/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0009.png", "metadata/rollout_136/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0010.png", "metadata/rollout_136/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0011.png", "metadata/rollout_136/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0012.png", "metadata/rollout_136/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0013.png", "metadata/rollout_136/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0014.png", "metadata/rollout_136/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0015.png", "metadata/rollout_136/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0016.png", "metadata/rollout_136/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0017.png", "metadata/rollout_136/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0018.png", "metadata/rollout_136/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0019.png", "metadata/rollout_136/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0020.png", "metadata/rollout_136/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_136/agentview/0021.png", "metadata/rollout_136/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0022.png", "metadata/rollout_136/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0023.png", "metadata/rollout_136/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0024.png", "metadata/rollout_136/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_136/agentview/0025.png", "metadata/rollout_136/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_136/agentview/0026.png", "metadata/rollout_136/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_136/agentview/0027.png", "metadata/rollout_136/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_136/agentview/0028.png", "metadata/rollout_136/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_136/agentview/0029.png", "metadata/rollout_136/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_136/agentview/0030.png", "metadata/rollout_136/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_136/agentview/0031.png", "metadata/rollout_136/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0032.png", "metadata/rollout_136/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_136/agentview/0033.png", "metadata/rollout_136/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0034.png", "metadata/rollout_136/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0035.png", "metadata/rollout_136/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_BACK, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0036.png", "metadata/rollout_136/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0037.png", "metadata/rollout_136/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0038.png", "metadata/rollout_136/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0039.png", "metadata/rollout_136/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0040.png", "metadata/rollout_136/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0041.png", "metadata/rollout_136/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0042.png", "metadata/rollout_136/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0043.png", "metadata/rollout_136/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0044.png", "metadata/rollout_136/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0045.png", "metadata/rollout_136/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0046.png", "metadata/rollout_136/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0047.png", "metadata/rollout_136/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0048.png", "metadata/rollout_136/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0049.png", "metadata/rollout_136/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0050.png", "metadata/rollout_136/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_136/agentview/0051.png", "metadata/rollout_136/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0052.png", "metadata/rollout_136/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0053.png", "metadata/rollout_136/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0054.png", "metadata/rollout_136/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_136/agentview/0055.png", "metadata/rollout_136/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_136/agentview/0056.png", "metadata/rollout_136/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_136/agentview/0057.png", "metadata/rollout_136/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_136/agentview/0058.png", "metadata/rollout_136/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_136/agentview/0059.png", "metadata/rollout_136/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_136/agentview/0060.png", "metadata/rollout_136/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_136/agentview/0060.png", "metadata/rollout_136/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0000.png", "metadata/rollout_137/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0001.png", "metadata/rollout_137/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0002.png", "metadata/rollout_137/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0003.png", "metadata/rollout_137/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0004.png", "metadata/rollout_137/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0005.png", "metadata/rollout_137/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0006.png", "metadata/rollout_137/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0007.png", "metadata/rollout_137/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0008.png", "metadata/rollout_137/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0009.png", "metadata/rollout_137/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0010.png", "metadata/rollout_137/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0011.png", "metadata/rollout_137/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0012.png", "metadata/rollout_137/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0013.png", "metadata/rollout_137/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_137/agentview/0014.png", "metadata/rollout_137/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_137/agentview/0015.png", "metadata/rollout_137/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_137/agentview/0016.png", "metadata/rollout_137/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_137/agentview/0017.png", "metadata/rollout_137/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_137/agentview/0018.png", "metadata/rollout_137/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_137/agentview/0019.png", "metadata/rollout_137/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_137/agentview/0020.png", "metadata/rollout_137/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0021.png", "metadata/rollout_137/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0022.png", "metadata/rollout_137/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0023.png", "metadata/rollout_137/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0024.png", "metadata/rollout_137/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0025.png", "metadata/rollout_137/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0026.png", "metadata/rollout_137/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0027.png", "metadata/rollout_137/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_137/agentview/0028.png", "metadata/rollout_137/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_137/agentview/0029.png", "metadata/rollout_137/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_137/agentview/0030.png", "metadata/rollout_137/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_137/agentview/0031.png", "metadata/rollout_137/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_137/agentview/0032.png", "metadata/rollout_137/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_137/agentview/0033.png", "metadata/rollout_137/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_137/agentview/0034.png", "metadata/rollout_137/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0035.png", "metadata/rollout_137/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0036.png", "metadata/rollout_137/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0037.png", "metadata/rollout_137/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0038.png", "metadata/rollout_137/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_DOWN, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_137/agentview/0039.png", "metadata/rollout_137/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0040.png", "metadata/rollout_137/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0041.png", "metadata/rollout_137/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0042.png", "metadata/rollout_137/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0043.png", "metadata/rollout_137/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0044.png", "metadata/rollout_137/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0045.png", "metadata/rollout_137/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0046.png", "metadata/rollout_137/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0047.png", "metadata/rollout_137/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0048.png", "metadata/rollout_137/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0049.png", "metadata/rollout_137/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_137/agentview/0050.png", "metadata/rollout_137/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0051.png", "metadata/rollout_137/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0052.png", "metadata/rollout_137/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0053.png", "metadata/rollout_137/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_137/agentview/0054.png", "metadata/rollout_137/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_137/agentview/0055.png", "metadata/rollout_137/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_137/agentview/0056.png", "metadata/rollout_137/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_137/agentview/0057.png", "metadata/rollout_137/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_137/agentview/0058.png", "metadata/rollout_137/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_137/agentview/0058.png", "metadata/rollout_137/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0000.png", "metadata/rollout_138/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0001.png", "metadata/rollout_138/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_138/agentview/0002.png", "metadata/rollout_138/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0003.png", "metadata/rollout_138/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0004.png", "metadata/rollout_138/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0005.png", "metadata/rollout_138/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0006.png", "metadata/rollout_138/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0007.png", "metadata/rollout_138/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0008.png", "metadata/rollout_138/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0009.png", "metadata/rollout_138/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0010.png", "metadata/rollout_138/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0011.png", "metadata/rollout_138/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0012.png", "metadata/rollout_138/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0013.png", "metadata/rollout_138/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0014.png", "metadata/rollout_138/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0015.png", "metadata/rollout_138/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0016.png", "metadata/rollout_138/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0017.png", "metadata/rollout_138/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0018.png", "metadata/rollout_138/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0019.png", "metadata/rollout_138/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0020.png", "metadata/rollout_138/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_138/agentview/0021.png", "metadata/rollout_138/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0022.png", "metadata/rollout_138/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0023.png", "metadata/rollout_138/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0024.png", "metadata/rollout_138/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0025.png", "metadata/rollout_138/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0026.png", "metadata/rollout_138/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0027.png", "metadata/rollout_138/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0028.png", "metadata/rollout_138/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_138/agentview/0029.png", "metadata/rollout_138/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0030.png", "metadata/rollout_138/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0031.png", "metadata/rollout_138/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_LEFT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0032.png", "metadata/rollout_138/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_138/agentview/0033.png", "metadata/rollout_138/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_138/agentview/0034.png", "metadata/rollout_138/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_138/agentview/0035.png", "metadata/rollout_138/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_138/agentview/0036.png", "metadata/rollout_138/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_138/agentview/0037.png", "metadata/rollout_138/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0038.png", "metadata/rollout_138/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0039.png", "metadata/rollout_138/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0040.png", "metadata/rollout_138/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_138/agentview/0041.png", "metadata/rollout_138/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_138/agentview/0042.png", "metadata/rollout_138/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_138/agentview/0043.png", "metadata/rollout_138/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0044.png", "metadata/rollout_138/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0045.png", "metadata/rollout_138/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0046.png", "metadata/rollout_138/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_138/agentview/0047.png", "metadata/rollout_138/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_138/agentview/0047.png", "metadata/rollout_138/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0000.png", "metadata/rollout_139/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0001.png", "metadata/rollout_139/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0002.png", "metadata/rollout_139/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0003.png", "metadata/rollout_139/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0004.png", "metadata/rollout_139/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0005.png", "metadata/rollout_139/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0006.png", "metadata/rollout_139/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0007.png", "metadata/rollout_139/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0008.png", "metadata/rollout_139/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_139/agentview/0009.png", "metadata/rollout_139/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_139/agentview/0010.png", "metadata/rollout_139/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_139/agentview/0011.png", "metadata/rollout_139/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_139/agentview/0012.png", "metadata/rollout_139/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0013.png", "metadata/rollout_139/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0014.png", "metadata/rollout_139/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0015.png", "metadata/rollout_139/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0016.png", "metadata/rollout_139/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0017.png", "metadata/rollout_139/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0018.png", "metadata/rollout_139/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0019.png", "metadata/rollout_139/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0020.png", "metadata/rollout_139/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0021.png", "metadata/rollout_139/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0022.png", "metadata/rollout_139/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0023.png", "metadata/rollout_139/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0024.png", "metadata/rollout_139/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0025.png", "metadata/rollout_139/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0026.png", "metadata/rollout_139/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0027.png", "metadata/rollout_139/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0028.png", "metadata/rollout_139/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0029.png", "metadata/rollout_139/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0030.png", "metadata/rollout_139/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0031.png", "metadata/rollout_139/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_139/agentview/0032.png", "metadata/rollout_139/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0033.png", "metadata/rollout_139/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0034.png", "metadata/rollout_139/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0035.png", "metadata/rollout_139/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0036.png", "metadata/rollout_139/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0037.png", "metadata/rollout_139/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0038.png", "metadata/rollout_139/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0039.png", "metadata/rollout_139/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_139/agentview/0040.png", "metadata/rollout_139/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0041.png", "metadata/rollout_139/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0042.png", "metadata/rollout_139/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_RIGHT, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0043.png", "metadata/rollout_139/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0044.png", "metadata/rollout_139/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0045.png", "metadata/rollout_139/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0046.png", "metadata/rollout_139/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0047.png", "metadata/rollout_139/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0048.png", "metadata/rollout_139/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_139/agentview/0049.png", "metadata/rollout_139/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_139/agentview/0050.png", "metadata/rollout_139/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_139/agentview/0051.png", "metadata/rollout_139/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_139/agentview/0052.png", "metadata/rollout_139/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_139/agentview/0053.png", "metadata/rollout_139/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_139/agentview/0054.png", "metadata/rollout_139/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0055.png", "metadata/rollout_139/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0056.png", "metadata/rollout_139/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_139/agentview/0057.png", "metadata/rollout_139/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0058.png", "metadata/rollout_139/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0059.png", "metadata/rollout_139/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0060.png", "metadata/rollout_139/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_139/agentview/0061.png", "metadata/rollout_139/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_139/agentview/0062.png", "metadata/rollout_139/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_139/agentview/0063.png", "metadata/rollout_139/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0064.png", "metadata/rollout_139/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0065.png", "metadata/rollout_139/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0066.png", "metadata/rollout_139/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_139/agentview/0067.png", "metadata/rollout_139/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_139/agentview/0068.png", "metadata/rollout_139/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_139/agentview/0068.png", "metadata/rollout_139/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0000.png", "metadata/rollout_140/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0001.png", "metadata/rollout_140/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0002.png", "metadata/rollout_140/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0003.png", "metadata/rollout_140/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0004.png", "metadata/rollout_140/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0005.png", "metadata/rollout_140/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0006.png", "metadata/rollout_140/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0007.png", "metadata/rollout_140/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0008.png", "metadata/rollout_140/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0009.png", "metadata/rollout_140/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0010.png", "metadata/rollout_140/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0011.png", "metadata/rollout_140/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0012.png", "metadata/rollout_140/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0013.png", "metadata/rollout_140/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0014.png", "metadata/rollout_140/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0015.png", "metadata/rollout_140/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0016.png", "metadata/rollout_140/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0017.png", "metadata/rollout_140/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0018.png", "metadata/rollout_140/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0019.png", "metadata/rollout_140/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0020.png", "metadata/rollout_140/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0021.png", "metadata/rollout_140/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0022.png", "metadata/rollout_140/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0023.png", "metadata/rollout_140/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0024.png", "metadata/rollout_140/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0025.png", "metadata/rollout_140/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_140/agentview/0026.png", "metadata/rollout_140/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0027.png", "metadata/rollout_140/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0028.png", "metadata/rollout_140/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0029.png", "metadata/rollout_140/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0030.png", "metadata/rollout_140/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_140/agentview/0031.png", "metadata/rollout_140/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_140/agentview/0032.png", "metadata/rollout_140/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_140/agentview/0033.png", "metadata/rollout_140/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_140/agentview/0034.png", "metadata/rollout_140/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_140/agentview/0035.png", "metadata/rollout_140/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_140/agentview/0036.png", "metadata/rollout_140/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0037.png", "metadata/rollout_140/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0038.png", "metadata/rollout_140/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0039.png", "metadata/rollout_140/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0040.png", "metadata/rollout_140/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0041.png", "metadata/rollout_140/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0042.png", "metadata/rollout_140/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_140/agentview/0043.png", "metadata/rollout_140/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0044.png", "metadata/rollout_140/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0045.png", "metadata/rollout_140/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0046.png", "metadata/rollout_140/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0047.png", "metadata/rollout_140/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0048.png", "metadata/rollout_140/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0049.png", "metadata/rollout_140/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0050.png", "metadata/rollout_140/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0051.png", "metadata/rollout_140/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0052.png", "metadata/rollout_140/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0053.png", "metadata/rollout_140/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0054.png", "metadata/rollout_140/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_140/agentview/0055.png", "metadata/rollout_140/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0056.png", "metadata/rollout_140/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0057.png", "metadata/rollout_140/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0058.png", "metadata/rollout_140/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_140/agentview/0059.png", "metadata/rollout_140/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_140/agentview/0060.png", "metadata/rollout_140/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_140/agentview/0061.png", "metadata/rollout_140/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_140/agentview/0062.png", "metadata/rollout_140/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_140/agentview/0063.png", "metadata/rollout_140/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_140/agentview/0064.png", "metadata/rollout_140/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_140/agentview/0065.png", "metadata/rollout_140/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_140/agentview/0065.png", "metadata/rollout_140/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0000.png", "metadata/rollout_141/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0001.png", "metadata/rollout_141/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0002.png", "metadata/rollout_141/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0003.png", "metadata/rollout_141/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_141/agentview/0004.png", "metadata/rollout_141/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_141/agentview/0005.png", "metadata/rollout_141/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_141/agentview/0006.png", "metadata/rollout_141/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_141/agentview/0007.png", "metadata/rollout_141/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0008.png", "metadata/rollout_141/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0009.png", "metadata/rollout_141/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0010.png", "metadata/rollout_141/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0011.png", "metadata/rollout_141/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_141/agentview/0012.png", "metadata/rollout_141/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_141/agentview/0013.png", "metadata/rollout_141/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0014.png", "metadata/rollout_141/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0015.png", "metadata/rollout_141/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0016.png", "metadata/rollout_141/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0017.png", "metadata/rollout_141/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0018.png", "metadata/rollout_141/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0019.png", "metadata/rollout_141/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0020.png", "metadata/rollout_141/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0021.png", "metadata/rollout_141/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0022.png", "metadata/rollout_141/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0023.png", "metadata/rollout_141/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0024.png", "metadata/rollout_141/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_141/agentview/0025.png", "metadata/rollout_141/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_141/agentview/0026.png", "metadata/rollout_141/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0027.png", "metadata/rollout_141/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0028.png", "metadata/rollout_141/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0029.png", "metadata/rollout_141/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0030.png", "metadata/rollout_141/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0031.png", "metadata/rollout_141/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0032.png", "metadata/rollout_141/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_141/agentview/0033.png", "metadata/rollout_141/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_141/agentview/0034.png", "metadata/rollout_141/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0035.png", "metadata/rollout_141/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0036.png", "metadata/rollout_141/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0037.png", "metadata/rollout_141/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0038.png", "metadata/rollout_141/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0039.png", "metadata/rollout_141/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0040.png", "metadata/rollout_141/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_141/agentview/0041.png", "metadata/rollout_141/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_141/agentview/0042.png", "metadata/rollout_141/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0043.png", "metadata/rollout_141/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0044.png", "metadata/rollout_141/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0045.png", "metadata/rollout_141/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_141/agentview/0046.png", "metadata/rollout_141/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_141/agentview/0047.png", "metadata/rollout_141/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_141/agentview/0048.png", "metadata/rollout_141/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0049.png", "metadata/rollout_141/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0050.png", "metadata/rollout_141/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0051.png", "metadata/rollout_141/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_141/agentview/0052.png", "metadata/rollout_141/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_141/agentview/0053.png", "metadata/rollout_141/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_141/agentview/0054.png", "metadata/rollout_141/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_141/agentview/0054.png", "metadata/rollout_141/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_142/agentview/0000.png", "metadata/rollout_142/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_142/agentview/0001.png", "metadata/rollout_142/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_142/agentview/0002.png", "metadata/rollout_142/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_142/agentview/0003.png", "metadata/rollout_142/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0004.png", "metadata/rollout_142/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0005.png", "metadata/rollout_142/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0006.png", "metadata/rollout_142/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0007.png", "metadata/rollout_142/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_142/agentview/0008.png", "metadata/rollout_142/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_142/agentview/0009.png", "metadata/rollout_142/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0010.png", "metadata/rollout_142/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0011.png", "metadata/rollout_142/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0012.png", "metadata/rollout_142/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0013.png", "metadata/rollout_142/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0014.png", "metadata/rollout_142/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0015.png", "metadata/rollout_142/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0016.png", "metadata/rollout_142/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_142/agentview/0017.png", "metadata/rollout_142/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_142/agentview/0018.png", "metadata/rollout_142/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_142/agentview/0019.png", "metadata/rollout_142/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_142/agentview/0020.png", "metadata/rollout_142/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_142/agentview/0021.png", "metadata/rollout_142/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_142/agentview/0022.png", "metadata/rollout_142/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_142/agentview/0023.png", "metadata/rollout_142/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_142/agentview/0024.png", "metadata/rollout_142/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0025.png", "metadata/rollout_142/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0026.png", "metadata/rollout_142/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_142/agentview/0027.png", "metadata/rollout_142/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_142/agentview/0028.png", "metadata/rollout_142/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_142/agentview/0029.png", "metadata/rollout_142/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0030.png", "metadata/rollout_142/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0031.png", "metadata/rollout_142/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0032.png", "metadata/rollout_142/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0033.png", "metadata/rollout_142/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0034.png", "metadata/rollout_142/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0035.png", "metadata/rollout_142/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0036.png", "metadata/rollout_142/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0037.png", "metadata/rollout_142/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0038.png", "metadata/rollout_142/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0039.png", "metadata/rollout_142/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0040.png", "metadata/rollout_142/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_142/agentview/0041.png", "metadata/rollout_142/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_142/agentview/0042.png", "metadata/rollout_142/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0043.png", "metadata/rollout_142/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0044.png", "metadata/rollout_142/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0045.png", "metadata/rollout_142/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_142/agentview/0046.png", "metadata/rollout_142/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_142/agentview/0047.png", "metadata/rollout_142/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_142/agentview/0048.png", "metadata/rollout_142/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_142/agentview/0049.png", "metadata/rollout_142/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_142/agentview/0050.png", "metadata/rollout_142/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_142/agentview/0051.png", "metadata/rollout_142/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_142/agentview/0051.png", "metadata/rollout_142/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0000.png", "metadata/rollout_143/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0001.png", "metadata/rollout_143/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0002.png", "metadata/rollout_143/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0003.png", "metadata/rollout_143/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0004.png", "metadata/rollout_143/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0005.png", "metadata/rollout_143/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0006.png", "metadata/rollout_143/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0007.png", "metadata/rollout_143/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0008.png", "metadata/rollout_143/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0009.png", "metadata/rollout_143/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0010.png", "metadata/rollout_143/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0011.png", "metadata/rollout_143/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_143/agentview/0012.png", "metadata/rollout_143/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_143/agentview/0013.png", "metadata/rollout_143/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_143/agentview/0014.png", "metadata/rollout_143/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_143/agentview/0015.png", "metadata/rollout_143/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_143/agentview/0016.png", "metadata/rollout_143/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_143/agentview/0017.png", "metadata/rollout_143/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0018.png", "metadata/rollout_143/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0019.png", "metadata/rollout_143/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0020.png", "metadata/rollout_143/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0021.png", "metadata/rollout_143/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0022.png", "metadata/rollout_143/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0023.png", "metadata/rollout_143/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0024.png", "metadata/rollout_143/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0025.png", "metadata/rollout_143/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0026.png", "metadata/rollout_143/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0027.png", "metadata/rollout_143/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0028.png", "metadata/rollout_143/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_143/agentview/0029.png", "metadata/rollout_143/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_143/agentview/0030.png", "metadata/rollout_143/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0031.png", "metadata/rollout_143/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0032.png", "metadata/rollout_143/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0033.png", "metadata/rollout_143/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0034.png", "metadata/rollout_143/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0035.png", "metadata/rollout_143/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0036.png", "metadata/rollout_143/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_143/agentview/0037.png", "metadata/rollout_143/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_143/agentview/0038.png", "metadata/rollout_143/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0039.png", "metadata/rollout_143/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0040.png", "metadata/rollout_143/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0041.png", "metadata/rollout_143/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0042.png", "metadata/rollout_143/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0043.png", "metadata/rollout_143/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0044.png", "metadata/rollout_143/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0045.png", "metadata/rollout_143/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0046.png", "metadata/rollout_143/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0047.png", "metadata/rollout_143/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_143/agentview/0048.png", "metadata/rollout_143/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_143/agentview/0049.png", "metadata/rollout_143/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_143/agentview/0050.png", "metadata/rollout_143/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_143/agentview/0051.png", "metadata/rollout_143/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_143/agentview/0052.png", "metadata/rollout_143/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_143/agentview/0053.png", "metadata/rollout_143/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_143/agentview/0054.png", "metadata/rollout_143/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0055.png", "metadata/rollout_143/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0056.png", "metadata/rollout_143/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_143/agentview/0057.png", "metadata/rollout_143/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0058.png", "metadata/rollout_143/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0059.png", "metadata/rollout_143/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0060.png", "metadata/rollout_143/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_143/agentview/0061.png", "metadata/rollout_143/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_143/agentview/0062.png", "metadata/rollout_143/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0063.png", "metadata/rollout_143/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0064.png", "metadata/rollout_143/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0065.png", "metadata/rollout_143/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_143/agentview/0066.png", "metadata/rollout_143/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_143/agentview/0066.png", "metadata/rollout_143/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0000.png", "metadata/rollout_144/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0001.png", "metadata/rollout_144/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0002.png", "metadata/rollout_144/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0003.png", "metadata/rollout_144/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0004.png", "metadata/rollout_144/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0005.png", "metadata/rollout_144/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_144/agentview/0006.png", "metadata/rollout_144/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_144/agentview/0007.png", "metadata/rollout_144/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_144/agentview/0008.png", "metadata/rollout_144/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0009.png", "metadata/rollout_144/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0010.png", "metadata/rollout_144/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0011.png", "metadata/rollout_144/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0012.png", "metadata/rollout_144/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0013.png", "metadata/rollout_144/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0014.png", "metadata/rollout_144/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0015.png", "metadata/rollout_144/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0016.png", "metadata/rollout_144/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0017.png", "metadata/rollout_144/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0018.png", "metadata/rollout_144/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0019.png", "metadata/rollout_144/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_144/agentview/0020.png", "metadata/rollout_144/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0021.png", "metadata/rollout_144/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0022.png", "metadata/rollout_144/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0023.png", "metadata/rollout_144/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_144/agentview/0024.png", "metadata/rollout_144/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_144/agentview/0025.png", "metadata/rollout_144/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_144/agentview/0026.png", "metadata/rollout_144/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_144/agentview/0027.png", "metadata/rollout_144/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_144/agentview/0028.png", "metadata/rollout_144/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_144/agentview/0029.png", "metadata/rollout_144/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_144/agentview/0030.png", "metadata/rollout_144/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_144/agentview/0031.png", "metadata/rollout_144/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_144/agentview/0032.png", "metadata/rollout_144/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_144/agentview/0033.png", "metadata/rollout_144/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_144/agentview/0034.png", "metadata/rollout_144/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_144/agentview/0035.png", "metadata/rollout_144/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_144/agentview/0036.png", "metadata/rollout_144/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_144/agentview/0037.png", "metadata/rollout_144/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_144/agentview/0038.png", "metadata/rollout_144/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_144/agentview/0039.png", "metadata/rollout_144/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_144/agentview/0040.png", "metadata/rollout_144/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_144/agentview/0041.png", "metadata/rollout_144/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_144/agentview/0042.png", "metadata/rollout_144/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_144/agentview/0043.png", "metadata/rollout_144/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_144/agentview/0044.png", "metadata/rollout_144/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0045.png", "metadata/rollout_144/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0046.png", "metadata/rollout_144/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0047.png", "metadata/rollout_144/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0048.png", "metadata/rollout_144/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_144/agentview/0049.png", "metadata/rollout_144/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_144/agentview/0050.png", "metadata/rollout_144/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_144/agentview/0051.png", "metadata/rollout_144/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_144/agentview/0052.png", "metadata/rollout_144/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_144/agentview/0053.png", "metadata/rollout_144/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_144/agentview/0054.png", "metadata/rollout_144/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_144/agentview/0054.png", "metadata/rollout_144/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_145/agentview/0000.png", "metadata/rollout_145/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_145/agentview/0001.png", "metadata/rollout_145/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_145/agentview/0002.png", "metadata/rollout_145/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_145/agentview/0003.png", "metadata/rollout_145/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_145/agentview/0004.png", "metadata/rollout_145/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_145/agentview/0005.png", "metadata/rollout_145/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_145/agentview/0006.png", "metadata/rollout_145/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_145/agentview/0007.png", "metadata/rollout_145/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_145/agentview/0008.png", "metadata/rollout_145/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_145/agentview/0009.png", "metadata/rollout_145/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_145/agentview/0010.png", "metadata/rollout_145/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_145/agentview/0011.png", "metadata/rollout_145/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_145/agentview/0012.png", "metadata/rollout_145/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0013.png", "metadata/rollout_145/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0014.png", "metadata/rollout_145/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0015.png", "metadata/rollout_145/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0016.png", "metadata/rollout_145/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0017.png", "metadata/rollout_145/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0018.png", "metadata/rollout_145/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0019.png", "metadata/rollout_145/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0020.png", "metadata/rollout_145/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_145/agentview/0021.png", "metadata/rollout_145/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_145/agentview/0022.png", "metadata/rollout_145/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0023.png", "metadata/rollout_145/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0024.png", "metadata/rollout_145/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0025.png", "metadata/rollout_145/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0026.png", "metadata/rollout_145/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_145/agentview/0027.png", "metadata/rollout_145/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0028.png", "metadata/rollout_145/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0029.png", "metadata/rollout_145/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0030.png", "metadata/rollout_145/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0031.png", "metadata/rollout_145/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0032.png", "metadata/rollout_145/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0033.png", "metadata/rollout_145/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0034.png", "metadata/rollout_145/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_145/agentview/0035.png", "metadata/rollout_145/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0036.png", "metadata/rollout_145/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0037.png", "metadata/rollout_145/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_145/agentview/0038.png", "metadata/rollout_145/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_145/agentview/0039.png", "metadata/rollout_145/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_145/agentview/0040.png", "metadata/rollout_145/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_145/agentview/0041.png", "metadata/rollout_145/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0042.png", "metadata/rollout_145/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0043.png", "metadata/rollout_145/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0044.png", "metadata/rollout_145/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0045.png", "metadata/rollout_145/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0046.png", "metadata/rollout_145/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0047.png", "metadata/rollout_145/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0048.png", "metadata/rollout_145/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0049.png", "metadata/rollout_145/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0050.png", "metadata/rollout_145/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0051.png", "metadata/rollout_145/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0052.png", "metadata/rollout_145/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0053.png", "metadata/rollout_145/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0054.png", "metadata/rollout_145/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_145/agentview/0055.png", "metadata/rollout_145/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_145/agentview/0056.png", "metadata/rollout_145/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_145/agentview/0057.png", "metadata/rollout_145/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0058.png", "metadata/rollout_145/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0059.png", "metadata/rollout_145/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0060.png", "metadata/rollout_145/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_145/agentview/0061.png", "metadata/rollout_145/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_145/agentview/0061.png", "metadata/rollout_145/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0000.png", "metadata/rollout_146/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0001.png", "metadata/rollout_146/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0002.png", "metadata/rollout_146/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0003.png", "metadata/rollout_146/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0004.png", "metadata/rollout_146/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0005.png", "metadata/rollout_146/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0006.png", "metadata/rollout_146/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0007.png", "metadata/rollout_146/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0008.png", "metadata/rollout_146/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0009.png", "metadata/rollout_146/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0010.png", "metadata/rollout_146/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0011.png", "metadata/rollout_146/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0012.png", "metadata/rollout_146/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0013.png", "metadata/rollout_146/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0014.png", "metadata/rollout_146/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0015.png", "metadata/rollout_146/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0016.png", "metadata/rollout_146/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0017.png", "metadata/rollout_146/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0018.png", "metadata/rollout_146/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0019.png", "metadata/rollout_146/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0020.png", "metadata/rollout_146/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0021.png", "metadata/rollout_146/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0022.png", "metadata/rollout_146/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0023.png", "metadata/rollout_146/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0024.png", "metadata/rollout_146/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0025.png", "metadata/rollout_146/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_146/agentview/0026.png", "metadata/rollout_146/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0027.png", "metadata/rollout_146/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0028.png", "metadata/rollout_146/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0029.png", "metadata/rollout_146/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_146/agentview/0030.png", "metadata/rollout_146/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0031.png", "metadata/rollout_146/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0032.png", "metadata/rollout_146/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0033.png", "metadata/rollout_146/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0034.png", "metadata/rollout_146/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0035.png", "metadata/rollout_146/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0036.png", "metadata/rollout_146/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0037.png", "metadata/rollout_146/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0038.png", "metadata/rollout_146/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0039.png", "metadata/rollout_146/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0040.png", "metadata/rollout_146/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0041.png", "metadata/rollout_146/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0042.png", "metadata/rollout_146/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0043.png", "metadata/rollout_146/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0044.png", "metadata/rollout_146/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0045.png", "metadata/rollout_146/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0046.png", "metadata/rollout_146/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0047.png", "metadata/rollout_146/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0048.png", "metadata/rollout_146/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0049.png", "metadata/rollout_146/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0050.png", "metadata/rollout_146/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0051.png", "metadata/rollout_146/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0052.png", "metadata/rollout_146/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0053.png", "metadata/rollout_146/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0054.png", "metadata/rollout_146/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0055.png", "metadata/rollout_146/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0056.png", "metadata/rollout_146/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0057.png", "metadata/rollout_146/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0058.png", "metadata/rollout_146/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0059.png", "metadata/rollout_146/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_146/agentview/0060.png", "metadata/rollout_146/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0061.png", "metadata/rollout_146/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0062.png", "metadata/rollout_146/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0063.png", "metadata/rollout_146/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_146/agentview/0064.png", "metadata/rollout_146/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_146/agentview/0065.png", "metadata/rollout_146/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_146/agentview/0066.png", "metadata/rollout_146/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_146/agentview/0067.png", "metadata/rollout_146/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0068.png", "metadata/rollout_146/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0069.png", "metadata/rollout_146/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0070.png", "metadata/rollout_146/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_146/agentview/0071.png", "metadata/rollout_146/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_146/agentview/0071.png", "metadata/rollout_146/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_147/agentview/0000.png", "metadata/rollout_147/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_147/agentview/0001.png", "metadata/rollout_147/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_147/agentview/0002.png", "metadata/rollout_147/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_147/agentview/0003.png", "metadata/rollout_147/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_147/agentview/0004.png", "metadata/rollout_147/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_147/agentview/0005.png", "metadata/rollout_147/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0006.png", "metadata/rollout_147/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0007.png", "metadata/rollout_147/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0008.png", "metadata/rollout_147/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_147/agentview/0009.png", "metadata/rollout_147/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_147/agentview/0010.png", "metadata/rollout_147/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0011.png", "metadata/rollout_147/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0012.png", "metadata/rollout_147/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_147/agentview/0013.png", "metadata/rollout_147/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0014.png", "metadata/rollout_147/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0015.png", "metadata/rollout_147/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0016.png", "metadata/rollout_147/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0017.png", "metadata/rollout_147/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0018.png", "metadata/rollout_147/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0019.png", "metadata/rollout_147/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_147/agentview/0020.png", "metadata/rollout_147/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_147/agentview/0021.png", "metadata/rollout_147/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_147/agentview/0022.png", "metadata/rollout_147/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_147/agentview/0023.png", "metadata/rollout_147/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_147/agentview/0024.png", "metadata/rollout_147/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_147/agentview/0025.png", "metadata/rollout_147/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_147/agentview/0026.png", "metadata/rollout_147/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_147/agentview/0027.png", "metadata/rollout_147/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_147/agentview/0028.png", "metadata/rollout_147/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_147/agentview/0029.png", "metadata/rollout_147/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0030.png", "metadata/rollout_147/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_147/agentview/0031.png", "metadata/rollout_147/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_147/agentview/0032.png", "metadata/rollout_147/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_147/agentview/0033.png", "metadata/rollout_147/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_147/agentview/0034.png", "metadata/rollout_147/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_147/agentview/0035.png", "metadata/rollout_147/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0036.png", "metadata/rollout_147/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0037.png", "metadata/rollout_147/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0038.png", "metadata/rollout_147/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_147/agentview/0039.png", "metadata/rollout_147/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_147/agentview/0040.png", "metadata/rollout_147/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_147/agentview/0041.png", "metadata/rollout_147/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_147/agentview/0042.png", "metadata/rollout_147/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_147/agentview/0043.png", "metadata/rollout_147/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_147/agentview/0044.png", "metadata/rollout_147/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_147/agentview/0044.png", "metadata/rollout_147/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0000.png", "metadata/rollout_148/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0001.png", "metadata/rollout_148/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0002.png", "metadata/rollout_148/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0003.png", "metadata/rollout_148/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0004.png", "metadata/rollout_148/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0005.png", "metadata/rollout_148/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0006.png", "metadata/rollout_148/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0007.png", "metadata/rollout_148/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0008.png", "metadata/rollout_148/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0009.png", "metadata/rollout_148/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_148/agentview/0010.png", "metadata/rollout_148/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_148/agentview/0011.png", "metadata/rollout_148/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_148/agentview/0012.png", "metadata/rollout_148/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_148/agentview/0013.png", "metadata/rollout_148/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0014.png", "metadata/rollout_148/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0015.png", "metadata/rollout_148/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0016.png", "metadata/rollout_148/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0017.png", "metadata/rollout_148/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_148/agentview/0018.png", "metadata/rollout_148/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0019.png", "metadata/rollout_148/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0020.png", "metadata/rollout_148/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0021.png", "metadata/rollout_148/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0022.png", "metadata/rollout_148/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_148/agentview/0023.png", "metadata/rollout_148/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0024.png", "metadata/rollout_148/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0025.png", "metadata/rollout_148/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0026.png", "metadata/rollout_148/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_148/agentview/0027.png", "metadata/rollout_148/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0028.png", "metadata/rollout_148/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0029.png", "metadata/rollout_148/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0030.png", "metadata/rollout_148/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0031.png", "metadata/rollout_148/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0032.png", "metadata/rollout_148/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0033.png", "metadata/rollout_148/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_148/agentview/0034.png", "metadata/rollout_148/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_148/agentview/0035.png", "metadata/rollout_148/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_148/agentview/0036.png", "metadata/rollout_148/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0037.png", "metadata/rollout_148/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_148/agentview/0038.png", "metadata/rollout_148/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_148/agentview/0039.png", "metadata/rollout_148/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_148/agentview/0040.png", "metadata/rollout_148/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_148/agentview/0041.png", "metadata/rollout_148/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_148/agentview/0042.png", "metadata/rollout_148/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_148/agentview/0043.png", "metadata/rollout_148/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_148/agentview/0044.png", "metadata/rollout_148/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0045.png", "metadata/rollout_148/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0046.png", "metadata/rollout_148/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0047.png", "metadata/rollout_148/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_148/agentview/0048.png", "metadata/rollout_148/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_148/agentview/0049.png", "metadata/rollout_148/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0050.png", "metadata/rollout_148/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0051.png", "metadata/rollout_148/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0052.png", "metadata/rollout_148/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_148/agentview/0053.png", "metadata/rollout_148/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_148/agentview/0053.png", "metadata/rollout_148/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0000.png", "metadata/rollout_149/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0001.png", "metadata/rollout_149/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0002.png", "metadata/rollout_149/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_149/agentview/0003.png", "metadata/rollout_149/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_149/agentview/0004.png", "metadata/rollout_149/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_149/agentview/0005.png", "metadata/rollout_149/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_149/agentview/0006.png", "metadata/rollout_149/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_149/agentview/0007.png", "metadata/rollout_149/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_149/agentview/0008.png", "metadata/rollout_149/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0009.png", "metadata/rollout_149/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0010.png", "metadata/rollout_149/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0011.png", "metadata/rollout_149/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0012.png", "metadata/rollout_149/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0013.png", "metadata/rollout_149/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0014.png", "metadata/rollout_149/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0015.png", "metadata/rollout_149/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0016.png", "metadata/rollout_149/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_149/agentview/0017.png", "metadata/rollout_149/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_149/agentview/0018.png", "metadata/rollout_149/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_149/agentview/0019.png", "metadata/rollout_149/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_149/agentview/0020.png", "metadata/rollout_149/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_149/agentview/0021.png", "metadata/rollout_149/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_149/agentview/0022.png", "metadata/rollout_149/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_149/agentview/0023.png", "metadata/rollout_149/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_149/agentview/0024.png", "metadata/rollout_149/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_149/agentview/0025.png", "metadata/rollout_149/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0026.png", "metadata/rollout_149/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_149/agentview/0027.png", "metadata/rollout_149/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_149/agentview/0028.png", "metadata/rollout_149/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_149/agentview/0029.png", "metadata/rollout_149/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_149/agentview/0030.png", "metadata/rollout_149/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_149/agentview/0031.png", "metadata/rollout_149/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_149/agentview/0032.png", "metadata/rollout_149/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0033.png", "metadata/rollout_149/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0034.png", "metadata/rollout_149/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0035.png", "metadata/rollout_149/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_149/agentview/0036.png", "metadata/rollout_149/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_149/agentview/0037.png", "metadata/rollout_149/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_149/agentview/0038.png", "metadata/rollout_149/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_149/agentview/0039.png", "metadata/rollout_149/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_149/agentview/0040.png", "metadata/rollout_149/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_149/agentview/0041.png", "metadata/rollout_149/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_149/agentview/0042.png", "metadata/rollout_149/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_149/agentview/0043.png", "metadata/rollout_149/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_149/agentview/0043.png", "metadata/rollout_149/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0000.png", "metadata/rollout_150/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0001.png", "metadata/rollout_150/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0002.png", "metadata/rollout_150/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0003.png", "metadata/rollout_150/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0004.png", "metadata/rollout_150/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0005.png", "metadata/rollout_150/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0006.png", "metadata/rollout_150/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0007.png", "metadata/rollout_150/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0008.png", "metadata/rollout_150/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_150/agentview/0009.png", "metadata/rollout_150/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_150/agentview/0010.png", "metadata/rollout_150/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_150/agentview/0011.png", "metadata/rollout_150/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_150/agentview/0012.png", "metadata/rollout_150/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_150/agentview/0013.png", "metadata/rollout_150/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_150/agentview/0014.png", "metadata/rollout_150/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_150/agentview/0015.png", "metadata/rollout_150/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_150/agentview/0016.png", "metadata/rollout_150/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0017.png", "metadata/rollout_150/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0018.png", "metadata/rollout_150/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0019.png", "metadata/rollout_150/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0020.png", "metadata/rollout_150/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0021.png", "metadata/rollout_150/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0022.png", "metadata/rollout_150/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0023.png", "metadata/rollout_150/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0024.png", "metadata/rollout_150/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0025.png", "metadata/rollout_150/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0026.png", "metadata/rollout_150/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0027.png", "metadata/rollout_150/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0028.png", "metadata/rollout_150/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0029.png", "metadata/rollout_150/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0030.png", "metadata/rollout_150/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_150/agentview/0031.png", "metadata/rollout_150/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_150/agentview/0032.png", "metadata/rollout_150/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0033.png", "metadata/rollout_150/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0034.png", "metadata/rollout_150/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0035.png", "metadata/rollout_150/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0036.png", "metadata/rollout_150/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0037.png", "metadata/rollout_150/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0038.png", "metadata/rollout_150/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0039.png", "metadata/rollout_150/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0040.png", "metadata/rollout_150/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_150/agentview/0041.png", "metadata/rollout_150/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_150/agentview/0042.png", "metadata/rollout_150/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0043.png", "metadata/rollout_150/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0044.png", "metadata/rollout_150/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0045.png", "metadata/rollout_150/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0046.png", "metadata/rollout_150/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0047.png", "metadata/rollout_150/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0048.png", "metadata/rollout_150/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_150/agentview/0049.png", "metadata/rollout_150/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_150/agentview/0050.png", "metadata/rollout_150/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_150/agentview/0051.png", "metadata/rollout_150/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_150/agentview/0052.png", "metadata/rollout_150/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_150/agentview/0053.png", "metadata/rollout_150/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0054.png", "metadata/rollout_150/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0055.png", "metadata/rollout_150/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0056.png", "metadata/rollout_150/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0057.png", "metadata/rollout_150/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0058.png", "metadata/rollout_150/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0059.png", "metadata/rollout_150/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0060.png", "metadata/rollout_150/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0061.png", "metadata/rollout_150/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_150/agentview/0062.png", "metadata/rollout_150/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_150/agentview/0063.png", "metadata/rollout_150/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_150/agentview/0064.png", "metadata/rollout_150/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0065.png", "metadata/rollout_150/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0066.png", "metadata/rollout_150/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0067.png", "metadata/rollout_150/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_150/agentview/0068.png", "metadata/rollout_150/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_150/agentview/0069.png", "metadata/rollout_150/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_150/agentview/0069.png", "metadata/rollout_150/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0000.png", "metadata/rollout_151/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0001.png", "metadata/rollout_151/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0002.png", "metadata/rollout_151/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0003.png", "metadata/rollout_151/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0004.png", "metadata/rollout_151/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_151/agentview/0005.png", "metadata/rollout_151/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_151/agentview/0006.png", "metadata/rollout_151/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_151/agentview/0007.png", "metadata/rollout_151/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_151/agentview/0008.png", "metadata/rollout_151/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_151/agentview/0009.png", "metadata/rollout_151/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_151/agentview/0010.png", "metadata/rollout_151/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_151/agentview/0011.png", "metadata/rollout_151/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_151/agentview/0012.png", "metadata/rollout_151/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_151/agentview/0013.png", "metadata/rollout_151/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0014.png", "metadata/rollout_151/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0015.png", "metadata/rollout_151/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0016.png", "metadata/rollout_151/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0017.png", "metadata/rollout_151/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0018.png", "metadata/rollout_151/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0019.png", "metadata/rollout_151/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0020.png", "metadata/rollout_151/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_151/agentview/0021.png", "metadata/rollout_151/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_151/agentview/0022.png", "metadata/rollout_151/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0023.png", "metadata/rollout_151/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0024.png", "metadata/rollout_151/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0025.png", "metadata/rollout_151/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0026.png", "metadata/rollout_151/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0027.png", "metadata/rollout_151/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0028.png", "metadata/rollout_151/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0029.png", "metadata/rollout_151/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0030.png", "metadata/rollout_151/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0031.png", "metadata/rollout_151/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0032.png", "metadata/rollout_151/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0033.png", "metadata/rollout_151/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0034.png", "metadata/rollout_151/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0035.png", "metadata/rollout_151/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0036.png", "metadata/rollout_151/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0037.png", "metadata/rollout_151/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0038.png", "metadata/rollout_151/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_151/agentview/0039.png", "metadata/rollout_151/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0040.png", "metadata/rollout_151/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0041.png", "metadata/rollout_151/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0042.png", "metadata/rollout_151/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_151/agentview/0043.png", "metadata/rollout_151/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_151/agentview/0044.png", "metadata/rollout_151/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0045.png", "metadata/rollout_151/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0046.png", "metadata/rollout_151/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0047.png", "metadata/rollout_151/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_151/agentview/0048.png", "metadata/rollout_151/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_151/agentview/0048.png", "metadata/rollout_151/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0000.png", "metadata/rollout_152/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0001.png", "metadata/rollout_152/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0002.png", "metadata/rollout_152/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0003.png", "metadata/rollout_152/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0004.png", "metadata/rollout_152/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0005.png", "metadata/rollout_152/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0006.png", "metadata/rollout_152/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0007.png", "metadata/rollout_152/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0008.png", "metadata/rollout_152/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0009.png", "metadata/rollout_152/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0010.png", "metadata/rollout_152/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0011.png", "metadata/rollout_152/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0012.png", "metadata/rollout_152/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0013.png", "metadata/rollout_152/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0014.png", "metadata/rollout_152/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0015.png", "metadata/rollout_152/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0016.png", "metadata/rollout_152/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0017.png", "metadata/rollout_152/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0018.png", "metadata/rollout_152/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0019.png", "metadata/rollout_152/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0020.png", "metadata/rollout_152/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0021.png", "metadata/rollout_152/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0022.png", "metadata/rollout_152/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0023.png", "metadata/rollout_152/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0024.png", "metadata/rollout_152/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0025.png", "metadata/rollout_152/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0026.png", "metadata/rollout_152/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_152/agentview/0027.png", "metadata/rollout_152/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_152/agentview/0028.png", "metadata/rollout_152/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0029.png", "metadata/rollout_152/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0030.png", "metadata/rollout_152/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0031.png", "metadata/rollout_152/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0032.png", "metadata/rollout_152/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_152/agentview/0033.png", "metadata/rollout_152/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0034.png", "metadata/rollout_152/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0035.png", "metadata/rollout_152/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0036.png", "metadata/rollout_152/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0037.png", "metadata/rollout_152/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0038.png", "metadata/rollout_152/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0039.png", "metadata/rollout_152/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_152/agentview/0040.png", "metadata/rollout_152/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_152/agentview/0041.png", "metadata/rollout_152/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0042.png", "metadata/rollout_152/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_152/agentview/0043.png", "metadata/rollout_152/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_BACK, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_152/agentview/0044.png", "metadata/rollout_152/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_152/agentview/0045.png", "metadata/rollout_152/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_152/agentview/0046.png", "metadata/rollout_152/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_152/agentview/0047.png", "metadata/rollout_152/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_152/agentview/0048.png", "metadata/rollout_152/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_152/agentview/0049.png", "metadata/rollout_152/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0050.png", "metadata/rollout_152/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0051.png", "metadata/rollout_152/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0052.png", "metadata/rollout_152/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_152/agentview/0053.png", "metadata/rollout_152/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_152/agentview/0054.png", "metadata/rollout_152/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0055.png", "metadata/rollout_152/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0056.png", "metadata/rollout_152/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0057.png", "metadata/rollout_152/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_152/agentview/0058.png", "metadata/rollout_152/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_152/agentview/0058.png", "metadata/rollout_152/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0000.png", "metadata/rollout_153/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0001.png", "metadata/rollout_153/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0002.png", "metadata/rollout_153/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0003.png", "metadata/rollout_153/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0004.png", "metadata/rollout_153/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0005.png", "metadata/rollout_153/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_153/agentview/0006.png", "metadata/rollout_153/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_153/agentview/0007.png", "metadata/rollout_153/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_153/agentview/0008.png", "metadata/rollout_153/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_153/agentview/0009.png", "metadata/rollout_153/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_153/agentview/0010.png", "metadata/rollout_153/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_153/agentview/0011.png", "metadata/rollout_153/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0012.png", "metadata/rollout_153/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0013.png", "metadata/rollout_153/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0014.png", "metadata/rollout_153/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0015.png", "metadata/rollout_153/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0016.png", "metadata/rollout_153/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0017.png", "metadata/rollout_153/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0018.png", "metadata/rollout_153/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0019.png", "metadata/rollout_153/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0020.png", "metadata/rollout_153/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0021.png", "metadata/rollout_153/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0022.png", "metadata/rollout_153/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0023.png", "metadata/rollout_153/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0024.png", "metadata/rollout_153/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0025.png", "metadata/rollout_153/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0026.png", "metadata/rollout_153/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0027.png", "metadata/rollout_153/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0028.png", "metadata/rollout_153/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_153/agentview/0029.png", "metadata/rollout_153/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0030.png", "metadata/rollout_153/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0031.png", "metadata/rollout_153/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0032.png", "metadata/rollout_153/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0033.png", "metadata/rollout_153/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0034.png", "metadata/rollout_153/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0035.png", "metadata/rollout_153/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0036.png", "metadata/rollout_153/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_153/agentview/0037.png", "metadata/rollout_153/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0038.png", "metadata/rollout_153/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0039.png", "metadata/rollout_153/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0040.png", "metadata/rollout_153/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_153/agentview/0041.png", "metadata/rollout_153/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0042.png", "metadata/rollout_153/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0043.png", "metadata/rollout_153/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0044.png", "metadata/rollout_153/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0045.png", "metadata/rollout_153/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0046.png", "metadata/rollout_153/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0047.png", "metadata/rollout_153/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0048.png", "metadata/rollout_153/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0049.png", "metadata/rollout_153/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0050.png", "metadata/rollout_153/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_153/agentview/0051.png", "metadata/rollout_153/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0052.png", "metadata/rollout_153/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0053.png", "metadata/rollout_153/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0054.png", "metadata/rollout_153/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_153/agentview/0055.png", "metadata/rollout_153/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_153/agentview/0056.png", "metadata/rollout_153/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_153/agentview/0057.png", "metadata/rollout_153/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0058.png", "metadata/rollout_153/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0059.png", "metadata/rollout_153/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0060.png", "metadata/rollout_153/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_153/agentview/0061.png", "metadata/rollout_153/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_153/agentview/0062.png", "metadata/rollout_153/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_153/agentview/0062.png", "metadata/rollout_153/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_154/agentview/0000.png", "metadata/rollout_154/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_154/agentview/0001.png", "metadata/rollout_154/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_154/agentview/0002.png", "metadata/rollout_154/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_154/agentview/0003.png", "metadata/rollout_154/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_154/agentview/0004.png", "metadata/rollout_154/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_154/agentview/0005.png", "metadata/rollout_154/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0006.png", "metadata/rollout_154/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0007.png", "metadata/rollout_154/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0008.png", "metadata/rollout_154/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_154/agentview/0009.png", "metadata/rollout_154/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_154/agentview/0010.png", "metadata/rollout_154/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_154/agentview/0011.png", "metadata/rollout_154/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_154/agentview/0012.png", "metadata/rollout_154/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_154/agentview/0013.png", "metadata/rollout_154/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0014.png", "metadata/rollout_154/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0015.png", "metadata/rollout_154/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0016.png", "metadata/rollout_154/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0017.png", "metadata/rollout_154/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0018.png", "metadata/rollout_154/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_154/agentview/0019.png", "metadata/rollout_154/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0020.png", "metadata/rollout_154/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0021.png", "metadata/rollout_154/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0022.png", "metadata/rollout_154/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0023.png", "metadata/rollout_154/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0024.png", "metadata/rollout_154/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0025.png", "metadata/rollout_154/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_154/agentview/0026.png", "metadata/rollout_154/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0027.png", "metadata/rollout_154/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0028.png", "metadata/rollout_154/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0029.png", "metadata/rollout_154/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0030.png", "metadata/rollout_154/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0031.png", "metadata/rollout_154/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0032.png", "metadata/rollout_154/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_154/agentview/0033.png", "metadata/rollout_154/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0034.png", "metadata/rollout_154/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_154/agentview/0035.png", "metadata/rollout_154/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_154/agentview/0036.png", "metadata/rollout_154/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_154/agentview/0037.png", "metadata/rollout_154/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_154/agentview/0038.png", "metadata/rollout_154/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_154/agentview/0039.png", "metadata/rollout_154/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_154/agentview/0040.png", "metadata/rollout_154/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_154/agentview/0041.png", "metadata/rollout_154/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_154/agentview/0042.png", "metadata/rollout_154/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_154/agentview/0043.png", "metadata/rollout_154/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_154/agentview/0044.png", "metadata/rollout_154/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0045.png", "metadata/rollout_154/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0046.png", "metadata/rollout_154/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0047.png", "metadata/rollout_154/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_154/agentview/0048.png", "metadata/rollout_154/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_154/agentview/0049.png", "metadata/rollout_154/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_154/agentview/0050.png", "metadata/rollout_154/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0051.png", "metadata/rollout_154/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0052.png", "metadata/rollout_154/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0053.png", "metadata/rollout_154/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_154/agentview/0054.png", "metadata/rollout_154/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_154/agentview/0055.png", "metadata/rollout_154/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_154/agentview/0055.png", "metadata/rollout_154/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0000.png", "metadata/rollout_155/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0001.png", "metadata/rollout_155/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0002.png", "metadata/rollout_155/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0003.png", "metadata/rollout_155/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0004.png", "metadata/rollout_155/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0005.png", "metadata/rollout_155/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0006.png", "metadata/rollout_155/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0007.png", "metadata/rollout_155/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0008.png", "metadata/rollout_155/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0009.png", "metadata/rollout_155/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0010.png", "metadata/rollout_155/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0011.png", "metadata/rollout_155/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0012.png", "metadata/rollout_155/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0013.png", "metadata/rollout_155/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0014.png", "metadata/rollout_155/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0015.png", "metadata/rollout_155/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0016.png", "metadata/rollout_155/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0017.png", "metadata/rollout_155/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0018.png", "metadata/rollout_155/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0019.png", "metadata/rollout_155/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0020.png", "metadata/rollout_155/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0021.png", "metadata/rollout_155/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0022.png", "metadata/rollout_155/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_155/agentview/0023.png", "metadata/rollout_155/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0024.png", "metadata/rollout_155/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0025.png", "metadata/rollout_155/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0026.png", "metadata/rollout_155/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0027.png", "metadata/rollout_155/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0028.png", "metadata/rollout_155/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0029.png", "metadata/rollout_155/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_155/agentview/0030.png", "metadata/rollout_155/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_155/agentview/0031.png", "metadata/rollout_155/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0032.png", "metadata/rollout_155/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0033.png", "metadata/rollout_155/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0034.png", "metadata/rollout_155/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0035.png", "metadata/rollout_155/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0036.png", "metadata/rollout_155/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0037.png", "metadata/rollout_155/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_155/agentview/0038.png", "metadata/rollout_155/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_155/agentview/0039.png", "metadata/rollout_155/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_155/agentview/0040.png", "metadata/rollout_155/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_155/agentview/0041.png", "metadata/rollout_155/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0042.png", "metadata/rollout_155/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0043.png", "metadata/rollout_155/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0044.png", "metadata/rollout_155/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_155/agentview/0045.png", "metadata/rollout_155/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_155/agentview/0046.png", "metadata/rollout_155/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0047.png", "metadata/rollout_155/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0048.png", "metadata/rollout_155/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0049.png", "metadata/rollout_155/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_155/agentview/0050.png", "metadata/rollout_155/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_155/agentview/0051.png", "metadata/rollout_155/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_155/agentview/0052.png", "metadata/rollout_155/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0053.png", "metadata/rollout_155/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0054.png", "metadata/rollout_155/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0055.png", "metadata/rollout_155/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_155/agentview/0056.png", "metadata/rollout_155/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_155/agentview/0057.png", "metadata/rollout_155/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_155/agentview/0057.png", "metadata/rollout_155/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_156/agentview/0000.png", "metadata/rollout_156/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0001.png", "metadata/rollout_156/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_156/agentview/0002.png", "metadata/rollout_156/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_156/agentview/0003.png", "metadata/rollout_156/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_156/agentview/0004.png", "metadata/rollout_156/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_156/agentview/0005.png", "metadata/rollout_156/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_156/agentview/0006.png", "metadata/rollout_156/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0007.png", "metadata/rollout_156/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0008.png", "metadata/rollout_156/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0009.png", "metadata/rollout_156/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0010.png", "metadata/rollout_156/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0011.png", "metadata/rollout_156/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0012.png", "metadata/rollout_156/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0013.png", "metadata/rollout_156/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0014.png", "metadata/rollout_156/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0015.png", "metadata/rollout_156/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0016.png", "metadata/rollout_156/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_156/agentview/0017.png", "metadata/rollout_156/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_156/agentview/0018.png", "metadata/rollout_156/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_156/agentview/0019.png", "metadata/rollout_156/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_156/agentview/0020.png", "metadata/rollout_156/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_156/agentview/0021.png", "metadata/rollout_156/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_156/agentview/0022.png", "metadata/rollout_156/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_156/agentview/0023.png", "metadata/rollout_156/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_156/agentview/0024.png", "metadata/rollout_156/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_156/agentview/0025.png", "metadata/rollout_156/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0026.png", "metadata/rollout_156/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_156/agentview/0027.png", "metadata/rollout_156/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_BACK, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_156/agentview/0028.png", "metadata/rollout_156/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_156/agentview/0029.png", "metadata/rollout_156/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_156/agentview/0030.png", "metadata/rollout_156/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_156/agentview/0031.png", "metadata/rollout_156/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_156/agentview/0032.png", "metadata/rollout_156/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_156/agentview/0033.png", "metadata/rollout_156/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0034.png", "metadata/rollout_156/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0035.png", "metadata/rollout_156/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0036.png", "metadata/rollout_156/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_156/agentview/0037.png", "metadata/rollout_156/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_156/agentview/0038.png", "metadata/rollout_156/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_156/agentview/0039.png", "metadata/rollout_156/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_156/agentview/0040.png", "metadata/rollout_156/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_156/agentview/0041.png", "metadata/rollout_156/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_156/agentview/0042.png", "metadata/rollout_156/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_156/agentview/0042.png", "metadata/rollout_156/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_157/agentview/0000.png", "metadata/rollout_157/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_157/agentview/0001.png", "metadata/rollout_157/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_157/agentview/0002.png", "metadata/rollout_157/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_157/agentview/0003.png", "metadata/rollout_157/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_157/agentview/0004.png", "metadata/rollout_157/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_157/agentview/0005.png", "metadata/rollout_157/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_157/agentview/0006.png", "metadata/rollout_157/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0007.png", "metadata/rollout_157/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0008.png", "metadata/rollout_157/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0009.png", "metadata/rollout_157/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0010.png", "metadata/rollout_157/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0011.png", "metadata/rollout_157/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0012.png", "metadata/rollout_157/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0013.png", "metadata/rollout_157/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0014.png", "metadata/rollout_157/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0015.png", "metadata/rollout_157/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0016.png", "metadata/rollout_157/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0017.png", "metadata/rollout_157/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_157/agentview/0018.png", "metadata/rollout_157/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_157/agentview/0019.png", "metadata/rollout_157/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_157/agentview/0020.png", "metadata/rollout_157/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_157/agentview/0021.png", "metadata/rollout_157/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_157/agentview/0022.png", "metadata/rollout_157/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_157/agentview/0023.png", "metadata/rollout_157/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_157/agentview/0024.png", "metadata/rollout_157/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0025.png", "metadata/rollout_157/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_157/agentview/0026.png", "metadata/rollout_157/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_157/agentview/0027.png", "metadata/rollout_157/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_157/agentview/0028.png", "metadata/rollout_157/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_157/agentview/0029.png", "metadata/rollout_157/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_157/agentview/0030.png", "metadata/rollout_157/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_157/agentview/0031.png", "metadata/rollout_157/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_157/agentview/0032.png", "metadata/rollout_157/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_157/agentview/0033.png", "metadata/rollout_157/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_157/agentview/0034.png", "metadata/rollout_157/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_157/agentview/0035.png", "metadata/rollout_157/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0036.png", "metadata/rollout_157/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0037.png", "metadata/rollout_157/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0038.png", "metadata/rollout_157/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_157/agentview/0039.png", "metadata/rollout_157/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_157/agentview/0040.png", "metadata/rollout_157/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_157/agentview/0041.png", "metadata/rollout_157/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_157/agentview/0042.png", "metadata/rollout_157/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_157/agentview/0043.png", "metadata/rollout_157/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_157/agentview/0044.png", "metadata/rollout_157/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_157/agentview/0045.png", "metadata/rollout_157/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_157/agentview/0046.png", "metadata/rollout_157/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_157/agentview/0046.png", "metadata/rollout_157/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_158/agentview/0000.png", "metadata/rollout_158/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_158/agentview/0001.png", "metadata/rollout_158/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_158/agentview/0002.png", "metadata/rollout_158/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_158/agentview/0003.png", "metadata/rollout_158/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0004.png", "metadata/rollout_158/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0005.png", "metadata/rollout_158/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0006.png", "metadata/rollout_158/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_158/agentview/0007.png", "metadata/rollout_158/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0008.png", "metadata/rollout_158/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0009.png", "metadata/rollout_158/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0010.png", "metadata/rollout_158/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0011.png", "metadata/rollout_158/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0012.png", "metadata/rollout_158/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0013.png", "metadata/rollout_158/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0014.png", "metadata/rollout_158/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0015.png", "metadata/rollout_158/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_158/agentview/0016.png", "metadata/rollout_158/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_158/agentview/0017.png", "metadata/rollout_158/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_158/agentview/0018.png", "metadata/rollout_158/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_158/agentview/0019.png", "metadata/rollout_158/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_158/agentview/0020.png", "metadata/rollout_158/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_158/agentview/0021.png", "metadata/rollout_158/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_158/agentview/0022.png", "metadata/rollout_158/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_158/agentview/0023.png", "metadata/rollout_158/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0024.png", "metadata/rollout_158/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_158/agentview/0025.png", "metadata/rollout_158/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_158/agentview/0026.png", "metadata/rollout_158/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_158/agentview/0027.png", "metadata/rollout_158/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_FWD, MV_DOWN, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_158/agentview/0028.png", "metadata/rollout_158/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_158/agentview/0029.png", "metadata/rollout_158/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_158/agentview/0030.png", "metadata/rollout_158/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_158/agentview/0031.png", "metadata/rollout_158/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_158/agentview/0032.png", "metadata/rollout_158/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_158/agentview/0033.png", "metadata/rollout_158/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_158/agentview/0034.png", "metadata/rollout_158/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_158/agentview/0035.png", "metadata/rollout_158/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_158/agentview/0036.png", "metadata/rollout_158/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0037.png", "metadata/rollout_158/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0038.png", "metadata/rollout_158/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0039.png", "metadata/rollout_158/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_158/agentview/0040.png", "metadata/rollout_158/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_158/agentview/0041.png", "metadata/rollout_158/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_158/agentview/0042.png", "metadata/rollout_158/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_158/agentview/0043.png", "metadata/rollout_158/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_158/agentview/0044.png", "metadata/rollout_158/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_158/agentview/0045.png", "metadata/rollout_158/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_158/agentview/0045.png", "metadata/rollout_158/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0000.png", "metadata/rollout_159/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0001.png", "metadata/rollout_159/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0002.png", "metadata/rollout_159/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0003.png", "metadata/rollout_159/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0004.png", "metadata/rollout_159/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0005.png", "metadata/rollout_159/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0006.png", "metadata/rollout_159/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0007.png", "metadata/rollout_159/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0008.png", "metadata/rollout_159/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0009.png", "metadata/rollout_159/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0010.png", "metadata/rollout_159/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0011.png", "metadata/rollout_159/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0012.png", "metadata/rollout_159/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0013.png", "metadata/rollout_159/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0014.png", "metadata/rollout_159/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0015.png", "metadata/rollout_159/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0016.png", "metadata/rollout_159/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0017.png", "metadata/rollout_159/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0018.png", "metadata/rollout_159/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0019.png", "metadata/rollout_159/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0020.png", "metadata/rollout_159/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0021.png", "metadata/rollout_159/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0022.png", "metadata/rollout_159/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0023.png", "metadata/rollout_159/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_159/agentview/0024.png", "metadata/rollout_159/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_159/agentview/0025.png", "metadata/rollout_159/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_159/agentview/0026.png", "metadata/rollout_159/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_159/agentview/0027.png", "metadata/rollout_159/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_159/agentview/0028.png", "metadata/rollout_159/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_159/agentview/0029.png", "metadata/rollout_159/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_159/agentview/0030.png", "metadata/rollout_159/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0031.png", "metadata/rollout_159/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0032.png", "metadata/rollout_159/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0033.png", "metadata/rollout_159/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0034.png", "metadata/rollout_159/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0035.png", "metadata/rollout_159/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0036.png", "metadata/rollout_159/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0037.png", "metadata/rollout_159/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0038.png", "metadata/rollout_159/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0039.png", "metadata/rollout_159/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0040.png", "metadata/rollout_159/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0041.png", "metadata/rollout_159/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0042.png", "metadata/rollout_159/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0043.png", "metadata/rollout_159/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0044.png", "metadata/rollout_159/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0045.png", "metadata/rollout_159/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0046.png", "metadata/rollout_159/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0047.png", "metadata/rollout_159/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0048.png", "metadata/rollout_159/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0049.png", "metadata/rollout_159/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0050.png", "metadata/rollout_159/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_159/agentview/0051.png", "metadata/rollout_159/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0052.png", "metadata/rollout_159/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0053.png", "metadata/rollout_159/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0054.png", "metadata/rollout_159/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_159/agentview/0055.png", "metadata/rollout_159/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_159/agentview/0056.png", "metadata/rollout_159/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_159/agentview/0057.png", "metadata/rollout_159/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_159/agentview/0058.png", "metadata/rollout_159/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_159/agentview/0059.png", "metadata/rollout_159/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_159/agentview/0060.png", "metadata/rollout_159/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_159/agentview/0061.png", "metadata/rollout_159/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_159/agentview/0062.png", "metadata/rollout_159/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_159/agentview/0062.png", "metadata/rollout_159/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0000.png", "metadata/rollout_160/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0001.png", "metadata/rollout_160/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_160/agentview/0002.png", "metadata/rollout_160/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_160/agentview/0003.png", "metadata/rollout_160/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0004.png", "metadata/rollout_160/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0005.png", "metadata/rollout_160/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0006.png", "metadata/rollout_160/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0007.png", "metadata/rollout_160/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_160/agentview/0008.png", "metadata/rollout_160/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_160/agentview/0009.png", "metadata/rollout_160/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0010.png", "metadata/rollout_160/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0011.png", "metadata/rollout_160/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0012.png", "metadata/rollout_160/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0013.png", "metadata/rollout_160/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0014.png", "metadata/rollout_160/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0015.png", "metadata/rollout_160/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0016.png", "metadata/rollout_160/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0017.png", "metadata/rollout_160/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_160/agentview/0018.png", "metadata/rollout_160/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_160/agentview/0019.png", "metadata/rollout_160/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_160/agentview/0020.png", "metadata/rollout_160/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_160/agentview/0021.png", "metadata/rollout_160/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_160/agentview/0022.png", "metadata/rollout_160/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_160/agentview/0023.png", "metadata/rollout_160/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0024.png", "metadata/rollout_160/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0025.png", "metadata/rollout_160/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0026.png", "metadata/rollout_160/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0027.png", "metadata/rollout_160/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0028.png", "metadata/rollout_160/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0029.png", "metadata/rollout_160/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0030.png", "metadata/rollout_160/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0031.png", "metadata/rollout_160/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0032.png", "metadata/rollout_160/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0033.png", "metadata/rollout_160/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0034.png", "metadata/rollout_160/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0035.png", "metadata/rollout_160/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0036.png", "metadata/rollout_160/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0037.png", "metadata/rollout_160/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0038.png", "metadata/rollout_160/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0039.png", "metadata/rollout_160/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0040.png", "metadata/rollout_160/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0041.png", "metadata/rollout_160/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0042.png", "metadata/rollout_160/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0043.png", "metadata/rollout_160/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0044.png", "metadata/rollout_160/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0045.png", "metadata/rollout_160/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0046.png", "metadata/rollout_160/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_160/agentview/0047.png", "metadata/rollout_160/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0048.png", "metadata/rollout_160/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0049.png", "metadata/rollout_160/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0050.png", "metadata/rollout_160/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_160/agentview/0051.png", "metadata/rollout_160/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_160/agentview/0052.png", "metadata/rollout_160/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_160/agentview/0053.png", "metadata/rollout_160/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_160/agentview/0054.png", "metadata/rollout_160/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_160/agentview/0055.png", "metadata/rollout_160/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_160/agentview/0056.png", "metadata/rollout_160/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_160/agentview/0057.png", "metadata/rollout_160/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_160/agentview/0057.png", "metadata/rollout_160/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0000.png", "metadata/rollout_161/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0001.png", "metadata/rollout_161/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0002.png", "metadata/rollout_161/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0003.png", "metadata/rollout_161/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0004.png", "metadata/rollout_161/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0005.png", "metadata/rollout_161/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_161/agentview/0006.png", "metadata/rollout_161/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_161/agentview/0007.png", "metadata/rollout_161/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0008.png", "metadata/rollout_161/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0009.png", "metadata/rollout_161/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0010.png", "metadata/rollout_161/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0011.png", "metadata/rollout_161/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0012.png", "metadata/rollout_161/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0013.png", "metadata/rollout_161/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0014.png", "metadata/rollout_161/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0015.png", "metadata/rollout_161/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0016.png", "metadata/rollout_161/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0017.png", "metadata/rollout_161/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0018.png", "metadata/rollout_161/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0019.png", "metadata/rollout_161/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0020.png", "metadata/rollout_161/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_161/agentview/0021.png", "metadata/rollout_161/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0022.png", "metadata/rollout_161/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0023.png", "metadata/rollout_161/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0024.png", "metadata/rollout_161/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0025.png", "metadata/rollout_161/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_161/agentview/0026.png", "metadata/rollout_161/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_161/agentview/0027.png", "metadata/rollout_161/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0028.png", "metadata/rollout_161/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0029.png", "metadata/rollout_161/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0030.png", "metadata/rollout_161/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0031.png", "metadata/rollout_161/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0032.png", "metadata/rollout_161/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0033.png", "metadata/rollout_161/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_161/agentview/0034.png", "metadata/rollout_161/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_161/agentview/0035.png", "metadata/rollout_161/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_161/agentview/0036.png", "metadata/rollout_161/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0037.png", "metadata/rollout_161/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_161/agentview/0038.png", "metadata/rollout_161/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_161/agentview/0039.png", "metadata/rollout_161/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_161/agentview/0040.png", "metadata/rollout_161/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_161/agentview/0041.png", "metadata/rollout_161/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_161/agentview/0042.png", "metadata/rollout_161/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_161/agentview/0043.png", "metadata/rollout_161/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_161/agentview/0044.png", "metadata/rollout_161/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0045.png", "metadata/rollout_161/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0046.png", "metadata/rollout_161/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0047.png", "metadata/rollout_161/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_161/agentview/0048.png", "metadata/rollout_161/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_161/agentview/0049.png", "metadata/rollout_161/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_161/agentview/0050.png", "metadata/rollout_161/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0051.png", "metadata/rollout_161/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0052.png", "metadata/rollout_161/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0053.png", "metadata/rollout_161/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_161/agentview/0054.png", "metadata/rollout_161/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_161/agentview/0054.png", "metadata/rollout_161/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0000.png", "metadata/rollout_162/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0001.png", "metadata/rollout_162/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0002.png", "metadata/rollout_162/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0003.png", "metadata/rollout_162/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0004.png", "metadata/rollout_162/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0005.png", "metadata/rollout_162/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0006.png", "metadata/rollout_162/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0007.png", "metadata/rollout_162/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0008.png", "metadata/rollout_162/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0009.png", "metadata/rollout_162/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0010.png", "metadata/rollout_162/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0011.png", "metadata/rollout_162/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0012.png", "metadata/rollout_162/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0013.png", "metadata/rollout_162/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0014.png", "metadata/rollout_162/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0015.png", "metadata/rollout_162/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0016.png", "metadata/rollout_162/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0017.png", "metadata/rollout_162/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0018.png", "metadata/rollout_162/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0019.png", "metadata/rollout_162/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0020.png", "metadata/rollout_162/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0021.png", "metadata/rollout_162/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0022.png", "metadata/rollout_162/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0023.png", "metadata/rollout_162/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0024.png", "metadata/rollout_162/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0025.png", "metadata/rollout_162/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0026.png", "metadata/rollout_162/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0027.png", "metadata/rollout_162/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_162/agentview/0028.png", "metadata/rollout_162/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0029.png", "metadata/rollout_162/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0030.png", "metadata/rollout_162/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_162/agentview/0031.png", "metadata/rollout_162/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0032.png", "metadata/rollout_162/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0033.png", "metadata/rollout_162/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0034.png", "metadata/rollout_162/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_162/agentview/0035.png", "metadata/rollout_162/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_162/agentview/0036.png", "metadata/rollout_162/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_162/agentview/0037.png", "metadata/rollout_162/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_162/agentview/0038.png", "metadata/rollout_162/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_162/agentview/0039.png", "metadata/rollout_162/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_162/agentview/0040.png", "metadata/rollout_162/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_162/agentview/0041.png", "metadata/rollout_162/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_162/agentview/0042.png", "metadata/rollout_162/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0043.png", "metadata/rollout_162/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0044.png", "metadata/rollout_162/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0045.png", "metadata/rollout_162/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0046.png", "metadata/rollout_162/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0047.png", "metadata/rollout_162/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0048.png", "metadata/rollout_162/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_162/agentview/0049.png", "metadata/rollout_162/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_162/agentview/0050.png", "metadata/rollout_162/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_162/agentview/0051.png", "metadata/rollout_162/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0052.png", "metadata/rollout_162/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0053.png", "metadata/rollout_162/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0054.png", "metadata/rollout_162/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0055.png", "metadata/rollout_162/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0056.png", "metadata/rollout_162/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0057.png", "metadata/rollout_162/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0058.png", "metadata/rollout_162/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0059.png", "metadata/rollout_162/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0060.png", "metadata/rollout_162/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0061.png", "metadata/rollout_162/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0062.png", "metadata/rollout_162/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0063.png", "metadata/rollout_162/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0064.png", "metadata/rollout_162/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_162/agentview/0065.png", "metadata/rollout_162/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0066.png", "metadata/rollout_162/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0067.png", "metadata/rollout_162/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0068.png", "metadata/rollout_162/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_162/agentview/0069.png", "metadata/rollout_162/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_162/agentview/0070.png", "metadata/rollout_162/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_162/agentview/0071.png", "metadata/rollout_162/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_162/agentview/0072.png", "metadata/rollout_162/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_162/agentview/0073.png", "metadata/rollout_162/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_162/agentview/0074.png", "metadata/rollout_162/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_162/agentview/0075.png", "metadata/rollout_162/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_162/agentview/0075.png", "metadata/rollout_162/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0000.png", "metadata/rollout_163/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0001.png", "metadata/rollout_163/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0002.png", "metadata/rollout_163/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0003.png", "metadata/rollout_163/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0004.png", "metadata/rollout_163/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0005.png", "metadata/rollout_163/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0006.png", "metadata/rollout_163/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0007.png", "metadata/rollout_163/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0008.png", "metadata/rollout_163/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0009.png", "metadata/rollout_163/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0010.png", "metadata/rollout_163/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_163/agentview/0011.png", "metadata/rollout_163/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_163/agentview/0012.png", "metadata/rollout_163/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_163/agentview/0013.png", "metadata/rollout_163/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_163/agentview/0014.png", "metadata/rollout_163/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_163/agentview/0015.png", "metadata/rollout_163/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0016.png", "metadata/rollout_163/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0017.png", "metadata/rollout_163/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0018.png", "metadata/rollout_163/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0019.png", "metadata/rollout_163/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0020.png", "metadata/rollout_163/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0021.png", "metadata/rollout_163/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0022.png", "metadata/rollout_163/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0023.png", "metadata/rollout_163/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0024.png", "metadata/rollout_163/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0025.png", "metadata/rollout_163/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_163/agentview/0026.png", "metadata/rollout_163/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0027.png", "metadata/rollout_163/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0028.png", "metadata/rollout_163/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0029.png", "metadata/rollout_163/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0030.png", "metadata/rollout_163/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0031.png", "metadata/rollout_163/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0032.png", "metadata/rollout_163/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0033.png", "metadata/rollout_163/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_163/agentview/0034.png", "metadata/rollout_163/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0035.png", "metadata/rollout_163/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_DOWN, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0036.png", "metadata/rollout_163/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_DOWN, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0037.png", "metadata/rollout_163/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_163/agentview/0038.png", "metadata/rollout_163/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_163/agentview/0039.png", "metadata/rollout_163/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0040.png", "metadata/rollout_163/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0041.png", "metadata/rollout_163/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0042.png", "metadata/rollout_163/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0043.png", "metadata/rollout_163/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0044.png", "metadata/rollout_163/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0045.png", "metadata/rollout_163/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0046.png", "metadata/rollout_163/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0047.png", "metadata/rollout_163/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0048.png", "metadata/rollout_163/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0049.png", "metadata/rollout_163/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0050.png", "metadata/rollout_163/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_163/agentview/0051.png", "metadata/rollout_163/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0052.png", "metadata/rollout_163/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0053.png", "metadata/rollout_163/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0054.png", "metadata/rollout_163/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_163/agentview/0055.png", "metadata/rollout_163/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_163/agentview/0056.png", "metadata/rollout_163/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_163/agentview/0057.png", "metadata/rollout_163/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0058.png", "metadata/rollout_163/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0059.png", "metadata/rollout_163/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0060.png", "metadata/rollout_163/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_163/agentview/0061.png", "metadata/rollout_163/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_163/agentview/0062.png", "metadata/rollout_163/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_163/agentview/0062.png", "metadata/rollout_163/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0000.png", "metadata/rollout_164/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0001.png", "metadata/rollout_164/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_164/agentview/0002.png", "metadata/rollout_164/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_164/agentview/0003.png", "metadata/rollout_164/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_164/agentview/0004.png", "metadata/rollout_164/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0005.png", "metadata/rollout_164/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0006.png", "metadata/rollout_164/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0007.png", "metadata/rollout_164/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0008.png", "metadata/rollout_164/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0009.png", "metadata/rollout_164/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0010.png", "metadata/rollout_164/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0011.png", "metadata/rollout_164/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0012.png", "metadata/rollout_164/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0013.png", "metadata/rollout_164/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0014.png", "metadata/rollout_164/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_164/agentview/0015.png", "metadata/rollout_164/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_164/agentview/0016.png", "metadata/rollout_164/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0017.png", "metadata/rollout_164/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0018.png", "metadata/rollout_164/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0019.png", "metadata/rollout_164/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0020.png", "metadata/rollout_164/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0021.png", "metadata/rollout_164/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0022.png", "metadata/rollout_164/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_164/agentview/0023.png", "metadata/rollout_164/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_164/agentview/0024.png", "metadata/rollout_164/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_164/agentview/0025.png", "metadata/rollout_164/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0026.png", "metadata/rollout_164/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_164/agentview/0027.png", "metadata/rollout_164/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_164/agentview/0028.png", "metadata/rollout_164/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_164/agentview/0029.png", "metadata/rollout_164/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_164/agentview/0030.png", "metadata/rollout_164/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_164/agentview/0031.png", "metadata/rollout_164/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_164/agentview/0032.png", "metadata/rollout_164/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_164/agentview/0033.png", "metadata/rollout_164/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_164/agentview/0034.png", "metadata/rollout_164/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0035.png", "metadata/rollout_164/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0036.png", "metadata/rollout_164/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0037.png", "metadata/rollout_164/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_164/agentview/0038.png", "metadata/rollout_164/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_164/agentview/0039.png", "metadata/rollout_164/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_164/agentview/0040.png", "metadata/rollout_164/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0041.png", "metadata/rollout_164/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0042.png", "metadata/rollout_164/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0043.png", "metadata/rollout_164/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_164/agentview/0044.png", "metadata/rollout_164/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_164/agentview/0044.png", "metadata/rollout_164/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0000.png", "metadata/rollout_165/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0001.png", "metadata/rollout_165/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0002.png", "metadata/rollout_165/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0003.png", "metadata/rollout_165/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0004.png", "metadata/rollout_165/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0005.png", "metadata/rollout_165/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0006.png", "metadata/rollout_165/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0007.png", "metadata/rollout_165/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0008.png", "metadata/rollout_165/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0009.png", "metadata/rollout_165/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0010.png", "metadata/rollout_165/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0011.png", "metadata/rollout_165/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_165/agentview/0012.png", "metadata/rollout_165/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_165/agentview/0013.png", "metadata/rollout_165/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_165/agentview/0014.png", "metadata/rollout_165/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_165/agentview/0015.png", "metadata/rollout_165/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_165/agentview/0016.png", "metadata/rollout_165/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0017.png", "metadata/rollout_165/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0018.png", "metadata/rollout_165/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0019.png", "metadata/rollout_165/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0020.png", "metadata/rollout_165/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0021.png", "metadata/rollout_165/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0022.png", "metadata/rollout_165/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_165/agentview/0023.png", "metadata/rollout_165/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_165/agentview/0024.png", "metadata/rollout_165/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0025.png", "metadata/rollout_165/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0026.png", "metadata/rollout_165/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0027.png", "metadata/rollout_165/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_165/agentview/0028.png", "metadata/rollout_165/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0029.png", "metadata/rollout_165/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0030.png", "metadata/rollout_165/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0031.png", "metadata/rollout_165/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0032.png", "metadata/rollout_165/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0033.png", "metadata/rollout_165/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0034.png", "metadata/rollout_165/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_165/agentview/0035.png", "metadata/rollout_165/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_165/agentview/0036.png", "metadata/rollout_165/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_165/agentview/0037.png", "metadata/rollout_165/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0038.png", "metadata/rollout_165/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_165/agentview/0039.png", "metadata/rollout_165/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_165/agentview/0040.png", "metadata/rollout_165/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_165/agentview/0041.png", "metadata/rollout_165/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_165/agentview/0042.png", "metadata/rollout_165/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_165/agentview/0043.png", "metadata/rollout_165/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_165/agentview/0044.png", "metadata/rollout_165/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_165/agentview/0045.png", "metadata/rollout_165/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0046.png", "metadata/rollout_165/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0047.png", "metadata/rollout_165/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0048.png", "metadata/rollout_165/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0049.png", "metadata/rollout_165/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_165/agentview/0050.png", "metadata/rollout_165/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_165/agentview/0051.png", "metadata/rollout_165/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0052.png", "metadata/rollout_165/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0053.png", "metadata/rollout_165/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0054.png", "metadata/rollout_165/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0055.png", "metadata/rollout_165/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_165/agentview/0056.png", "metadata/rollout_165/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_165/agentview/0056.png", "metadata/rollout_165/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0000.png", "metadata/rollout_166/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0001.png", "metadata/rollout_166/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0002.png", "metadata/rollout_166/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0003.png", "metadata/rollout_166/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0004.png", "metadata/rollout_166/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0005.png", "metadata/rollout_166/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0006.png", "metadata/rollout_166/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0007.png", "metadata/rollout_166/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0008.png", "metadata/rollout_166/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0009.png", "metadata/rollout_166/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0010.png", "metadata/rollout_166/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0011.png", "metadata/rollout_166/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0012.png", "metadata/rollout_166/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0013.png", "metadata/rollout_166/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0014.png", "metadata/rollout_166/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0015.png", "metadata/rollout_166/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0016.png", "metadata/rollout_166/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0017.png", "metadata/rollout_166/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0018.png", "metadata/rollout_166/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0019.png", "metadata/rollout_166/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0020.png", "metadata/rollout_166/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0021.png", "metadata/rollout_166/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0022.png", "metadata/rollout_166/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0023.png", "metadata/rollout_166/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0024.png", "metadata/rollout_166/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0025.png", "metadata/rollout_166/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0026.png", "metadata/rollout_166/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0027.png", "metadata/rollout_166/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_166/agentview/0028.png", "metadata/rollout_166/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_166/agentview/0029.png", "metadata/rollout_166/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0030.png", "metadata/rollout_166/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0031.png", "metadata/rollout_166/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_166/agentview/0032.png", "metadata/rollout_166/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_166/agentview/0033.png", "metadata/rollout_166/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_166/agentview/0034.png", "metadata/rollout_166/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_166/agentview/0035.png", "metadata/rollout_166/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_166/agentview/0036.png", "metadata/rollout_166/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_166/agentview/0037.png", "metadata/rollout_166/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0038.png", "metadata/rollout_166/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0039.png", "metadata/rollout_166/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_166/agentview/0040.png", "metadata/rollout_166/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_166/agentview/0041.png", "metadata/rollout_166/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0042.png", "metadata/rollout_166/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0043.png", "metadata/rollout_166/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_166/agentview/0044.png", "metadata/rollout_166/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0045.png", "metadata/rollout_166/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0046.png", "metadata/rollout_166/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0047.png", "metadata/rollout_166/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0048.png", "metadata/rollout_166/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0049.png", "metadata/rollout_166/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0050.png", "metadata/rollout_166/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0051.png", "metadata/rollout_166/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0052.png", "metadata/rollout_166/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0053.png", "metadata/rollout_166/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0054.png", "metadata/rollout_166/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_166/agentview/0055.png", "metadata/rollout_166/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0056.png", "metadata/rollout_166/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0057.png", "metadata/rollout_166/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0058.png", "metadata/rollout_166/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_166/agentview/0059.png", "metadata/rollout_166/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_166/agentview/0060.png", "metadata/rollout_166/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_166/agentview/0061.png", "metadata/rollout_166/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_166/agentview/0062.png", "metadata/rollout_166/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_166/agentview/0063.png", "metadata/rollout_166/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_166/agentview/0064.png", "metadata/rollout_166/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_166/agentview/0064.png", "metadata/rollout_166/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0000.png", "metadata/rollout_167/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0001.png", "metadata/rollout_167/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0002.png", "metadata/rollout_167/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0003.png", "metadata/rollout_167/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0004.png", "metadata/rollout_167/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0005.png", "metadata/rollout_167/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_167/agentview/0006.png", "metadata/rollout_167/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_167/agentview/0007.png", "metadata/rollout_167/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_167/agentview/0008.png", "metadata/rollout_167/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_167/agentview/0009.png", "metadata/rollout_167/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_167/agentview/0010.png", "metadata/rollout_167/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_167/agentview/0011.png", "metadata/rollout_167/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_167/agentview/0012.png", "metadata/rollout_167/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_167/agentview/0013.png", "metadata/rollout_167/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0014.png", "metadata/rollout_167/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0015.png", "metadata/rollout_167/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0016.png", "metadata/rollout_167/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0017.png", "metadata/rollout_167/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0018.png", "metadata/rollout_167/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_167/agentview/0019.png", "metadata/rollout_167/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_167/agentview/0020.png", "metadata/rollout_167/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_167/agentview/0021.png", "metadata/rollout_167/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_167/agentview/0022.png", "metadata/rollout_167/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_167/agentview/0023.png", "metadata/rollout_167/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_167/agentview/0024.png", "metadata/rollout_167/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_167/agentview/0025.png", "metadata/rollout_167/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0026.png", "metadata/rollout_167/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0027.png", "metadata/rollout_167/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_167/agentview/0028.png", "metadata/rollout_167/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_167/agentview/0029.png", "metadata/rollout_167/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_167/agentview/0030.png", "metadata/rollout_167/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_167/agentview/0031.png", "metadata/rollout_167/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_167/agentview/0032.png", "metadata/rollout_167/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_167/agentview/0033.png", "metadata/rollout_167/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_167/agentview/0034.png", "metadata/rollout_167/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0035.png", "metadata/rollout_167/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0036.png", "metadata/rollout_167/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0037.png", "metadata/rollout_167/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0038.png", "metadata/rollout_167/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0039.png", "metadata/rollout_167/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0040.png", "metadata/rollout_167/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0041.png", "metadata/rollout_167/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0042.png", "metadata/rollout_167/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0043.png", "metadata/rollout_167/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0044.png", "metadata/rollout_167/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0045.png", "metadata/rollout_167/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0046.png", "metadata/rollout_167/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0047.png", "metadata/rollout_167/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0048.png", "metadata/rollout_167/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0049.png", "metadata/rollout_167/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_167/agentview/0050.png", "metadata/rollout_167/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0051.png", "metadata/rollout_167/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0052.png", "metadata/rollout_167/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0053.png", "metadata/rollout_167/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_167/agentview/0054.png", "metadata/rollout_167/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_167/agentview/0055.png", "metadata/rollout_167/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_167/agentview/0056.png", "metadata/rollout_167/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_167/agentview/0057.png", "metadata/rollout_167/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_167/agentview/0058.png", "metadata/rollout_167/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_167/agentview/0059.png", "metadata/rollout_167/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_167/agentview/0060.png", "metadata/rollout_167/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_167/agentview/0060.png", "metadata/rollout_167/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0000.png", "metadata/rollout_168/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0001.png", "metadata/rollout_168/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0002.png", "metadata/rollout_168/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0003.png", "metadata/rollout_168/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0004.png", "metadata/rollout_168/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0005.png", "metadata/rollout_168/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0006.png", "metadata/rollout_168/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0007.png", "metadata/rollout_168/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0008.png", "metadata/rollout_168/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_168/agentview/0009.png", "metadata/rollout_168/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_168/agentview/0010.png", "metadata/rollout_168/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_168/agentview/0011.png", "metadata/rollout_168/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_168/agentview/0012.png", "metadata/rollout_168/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_168/agentview/0013.png", "metadata/rollout_168/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_168/agentview/0014.png", "metadata/rollout_168/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_168/agentview/0015.png", "metadata/rollout_168/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_168/agentview/0016.png", "metadata/rollout_168/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0017.png", "metadata/rollout_168/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0018.png", "metadata/rollout_168/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0019.png", "metadata/rollout_168/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0020.png", "metadata/rollout_168/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0021.png", "metadata/rollout_168/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0022.png", "metadata/rollout_168/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0023.png", "metadata/rollout_168/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0024.png", "metadata/rollout_168/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0025.png", "metadata/rollout_168/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_168/agentview/0026.png", "metadata/rollout_168/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0027.png", "metadata/rollout_168/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0028.png", "metadata/rollout_168/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0029.png", "metadata/rollout_168/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_168/agentview/0030.png", "metadata/rollout_168/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_168/agentview/0031.png", "metadata/rollout_168/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_168/agentview/0032.png", "metadata/rollout_168/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_168/agentview/0033.png", "metadata/rollout_168/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_168/agentview/0034.png", "metadata/rollout_168/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_168/agentview/0035.png", "metadata/rollout_168/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_168/agentview/0036.png", "metadata/rollout_168/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_168/agentview/0037.png", "metadata/rollout_168/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_168/agentview/0038.png", "metadata/rollout_168/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0039.png", "metadata/rollout_168/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0040.png", "metadata/rollout_168/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0041.png", "metadata/rollout_168/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0042.png", "metadata/rollout_168/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_168/agentview/0043.png", "metadata/rollout_168/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0044.png", "metadata/rollout_168/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0045.png", "metadata/rollout_168/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0046.png", "metadata/rollout_168/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0047.png", "metadata/rollout_168/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0048.png", "metadata/rollout_168/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0049.png", "metadata/rollout_168/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0050.png", "metadata/rollout_168/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0051.png", "metadata/rollout_168/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0052.png", "metadata/rollout_168/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0053.png", "metadata/rollout_168/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0054.png", "metadata/rollout_168/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0055.png", "metadata/rollout_168/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0056.png", "metadata/rollout_168/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0057.png", "metadata/rollout_168/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0058.png", "metadata/rollout_168/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_168/agentview/0059.png", "metadata/rollout_168/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0060.png", "metadata/rollout_168/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0061.png", "metadata/rollout_168/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0062.png", "metadata/rollout_168/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_168/agentview/0063.png", "metadata/rollout_168/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_168/agentview/0064.png", "metadata/rollout_168/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_168/agentview/0065.png", "metadata/rollout_168/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_168/agentview/0066.png", "metadata/rollout_168/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_168/agentview/0067.png", "metadata/rollout_168/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_168/agentview/0068.png", "metadata/rollout_168/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_168/agentview/0068.png", "metadata/rollout_168/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0000.png", "metadata/rollout_169/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0001.png", "metadata/rollout_169/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0002.png", "metadata/rollout_169/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_169/agentview/0003.png", "metadata/rollout_169/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_169/agentview/0004.png", "metadata/rollout_169/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_169/agentview/0005.png", "metadata/rollout_169/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0006.png", "metadata/rollout_169/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0007.png", "metadata/rollout_169/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0008.png", "metadata/rollout_169/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0009.png", "metadata/rollout_169/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0010.png", "metadata/rollout_169/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_169/agentview/0011.png", "metadata/rollout_169/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_169/agentview/0012.png", "metadata/rollout_169/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0013.png", "metadata/rollout_169/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0014.png", "metadata/rollout_169/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0015.png", "metadata/rollout_169/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0016.png", "metadata/rollout_169/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0017.png", "metadata/rollout_169/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0018.png", "metadata/rollout_169/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_169/agentview/0019.png", "metadata/rollout_169/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_169/agentview/0020.png", "metadata/rollout_169/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_169/agentview/0021.png", "metadata/rollout_169/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_169/agentview/0022.png", "metadata/rollout_169/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_169/agentview/0023.png", "metadata/rollout_169/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_169/agentview/0024.png", "metadata/rollout_169/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_169/agentview/0025.png", "metadata/rollout_169/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0026.png", "metadata/rollout_169/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0027.png", "metadata/rollout_169/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0028.png", "metadata/rollout_169/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_169/agentview/0029.png", "metadata/rollout_169/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0030.png", "metadata/rollout_169/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0031.png", "metadata/rollout_169/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0032.png", "metadata/rollout_169/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0033.png", "metadata/rollout_169/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0034.png", "metadata/rollout_169/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_169/agentview/0035.png", "metadata/rollout_169/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_169/agentview/0036.png", "metadata/rollout_169/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_169/agentview/0037.png", "metadata/rollout_169/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_169/agentview/0038.png", "metadata/rollout_169/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_169/agentview/0039.png", "metadata/rollout_169/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0040.png", "metadata/rollout_169/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0041.png", "metadata/rollout_169/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0042.png", "metadata/rollout_169/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_169/agentview/0043.png", "metadata/rollout_169/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0044.png", "metadata/rollout_169/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_169/agentview/0045.png", "metadata/rollout_169/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_169/agentview/0046.png", "metadata/rollout_169/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_169/agentview/0047.png", "metadata/rollout_169/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_169/agentview/0048.png", "metadata/rollout_169/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_169/agentview/0049.png", "metadata/rollout_169/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_169/agentview/0050.png", "metadata/rollout_169/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_169/agentview/0050.png", "metadata/rollout_169/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0000.png", "metadata/rollout_170/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0001.png", "metadata/rollout_170/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0002.png", "metadata/rollout_170/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0003.png", "metadata/rollout_170/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_170/agentview/0004.png", "metadata/rollout_170/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0005.png", "metadata/rollout_170/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0006.png", "metadata/rollout_170/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0007.png", "metadata/rollout_170/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0008.png", "metadata/rollout_170/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0009.png", "metadata/rollout_170/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0010.png", "metadata/rollout_170/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0011.png", "metadata/rollout_170/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0012.png", "metadata/rollout_170/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0013.png", "metadata/rollout_170/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0014.png", "metadata/rollout_170/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0015.png", "metadata/rollout_170/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0016.png", "metadata/rollout_170/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0017.png", "metadata/rollout_170/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0018.png", "metadata/rollout_170/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0019.png", "metadata/rollout_170/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0020.png", "metadata/rollout_170/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_170/agentview/0021.png", "metadata/rollout_170/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0022.png", "metadata/rollout_170/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0023.png", "metadata/rollout_170/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0024.png", "metadata/rollout_170/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_170/agentview/0025.png", "metadata/rollout_170/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0026.png", "metadata/rollout_170/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0027.png", "metadata/rollout_170/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0028.png", "metadata/rollout_170/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0029.png", "metadata/rollout_170/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0030.png", "metadata/rollout_170/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0031.png", "metadata/rollout_170/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_170/agentview/0032.png", "metadata/rollout_170/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_170/agentview/0033.png", "metadata/rollout_170/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_170/agentview/0034.png", "metadata/rollout_170/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_170/agentview/0035.png", "metadata/rollout_170/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0036.png", "metadata/rollout_170/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_170/agentview/0037.png", "metadata/rollout_170/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_170/agentview/0038.png", "metadata/rollout_170/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_170/agentview/0039.png", "metadata/rollout_170/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_170/agentview/0040.png", "metadata/rollout_170/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_170/agentview/0041.png", "metadata/rollout_170/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_170/agentview/0042.png", "metadata/rollout_170/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_170/agentview/0043.png", "metadata/rollout_170/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_170/agentview/0044.png", "metadata/rollout_170/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_170/agentview/0045.png", "metadata/rollout_170/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_170/agentview/0046.png", "metadata/rollout_170/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_170/agentview/0047.png", "metadata/rollout_170/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0048.png", "metadata/rollout_170/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0049.png", "metadata/rollout_170/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0050.png", "metadata/rollout_170/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_170/agentview/0051.png", "metadata/rollout_170/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_170/agentview/0052.png", "metadata/rollout_170/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0053.png", "metadata/rollout_170/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0054.png", "metadata/rollout_170/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0055.png", "metadata/rollout_170/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_170/agentview/0056.png", "metadata/rollout_170/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_170/agentview/0056.png", "metadata/rollout_170/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0000.png", "metadata/rollout_171/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0001.png", "metadata/rollout_171/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0002.png", "metadata/rollout_171/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0003.png", "metadata/rollout_171/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0004.png", "metadata/rollout_171/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0005.png", "metadata/rollout_171/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0006.png", "metadata/rollout_171/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0007.png", "metadata/rollout_171/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0008.png", "metadata/rollout_171/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0009.png", "metadata/rollout_171/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0010.png", "metadata/rollout_171/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0011.png", "metadata/rollout_171/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0012.png", "metadata/rollout_171/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0013.png", "metadata/rollout_171/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0014.png", "metadata/rollout_171/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0015.png", "metadata/rollout_171/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_171/agentview/0016.png", "metadata/rollout_171/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_171/agentview/0017.png", "metadata/rollout_171/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_171/agentview/0018.png", "metadata/rollout_171/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_171/agentview/0019.png", "metadata/rollout_171/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_171/agentview/0020.png", "metadata/rollout_171/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_171/agentview/0021.png", "metadata/rollout_171/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_171/agentview/0022.png", "metadata/rollout_171/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0023.png", "metadata/rollout_171/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0024.png", "metadata/rollout_171/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0025.png", "metadata/rollout_171/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0026.png", "metadata/rollout_171/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_171/agentview/0027.png", "metadata/rollout_171/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0028.png", "metadata/rollout_171/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0029.png", "metadata/rollout_171/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0030.png", "metadata/rollout_171/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0031.png", "metadata/rollout_171/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0032.png", "metadata/rollout_171/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_171/agentview/0033.png", "metadata/rollout_171/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0034.png", "metadata/rollout_171/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0035.png", "metadata/rollout_171/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0036.png", "metadata/rollout_171/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0037.png", "metadata/rollout_171/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0038.png", "metadata/rollout_171/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0039.png", "metadata/rollout_171/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0040.png", "metadata/rollout_171/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0041.png", "metadata/rollout_171/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_171/agentview/0042.png", "metadata/rollout_171/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_171/agentview/0043.png", "metadata/rollout_171/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0044.png", "metadata/rollout_171/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0045.png", "metadata/rollout_171/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0046.png", "metadata/rollout_171/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0047.png", "metadata/rollout_171/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0048.png", "metadata/rollout_171/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_171/agentview/0049.png", "metadata/rollout_171/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_171/agentview/0050.png", "metadata/rollout_171/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_171/agentview/0051.png", "metadata/rollout_171/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_171/agentview/0052.png", "metadata/rollout_171/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0053.png", "metadata/rollout_171/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0054.png", "metadata/rollout_171/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0055.png", "metadata/rollout_171/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0056.png", "metadata/rollout_171/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0057.png", "metadata/rollout_171/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_171/agentview/0058.png", "metadata/rollout_171/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0059.png", "metadata/rollout_171/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_171/agentview/0060.png", "metadata/rollout_171/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_171/agentview/0061.png", "metadata/rollout_171/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0062.png", "metadata/rollout_171/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0063.png", "metadata/rollout_171/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0064.png", "metadata/rollout_171/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_171/agentview/0065.png", "metadata/rollout_171/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_171/agentview/0066.png", "metadata/rollout_171/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_171/agentview/0066.png", "metadata/rollout_171/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0000.png", "metadata/rollout_172/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0001.png", "metadata/rollout_172/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0002.png", "metadata/rollout_172/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0003.png", "metadata/rollout_172/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0004.png", "metadata/rollout_172/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0005.png", "metadata/rollout_172/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0006.png", "metadata/rollout_172/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0007.png", "metadata/rollout_172/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0008.png", "metadata/rollout_172/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0009.png", "metadata/rollout_172/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0010.png", "metadata/rollout_172/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0011.png", "metadata/rollout_172/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_172/agentview/0012.png", "metadata/rollout_172/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_172/agentview/0013.png", "metadata/rollout_172/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_172/agentview/0014.png", "metadata/rollout_172/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_172/agentview/0015.png", "metadata/rollout_172/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_172/agentview/0016.png", "metadata/rollout_172/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_172/agentview/0017.png", "metadata/rollout_172/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0018.png", "metadata/rollout_172/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0019.png", "metadata/rollout_172/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0020.png", "metadata/rollout_172/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0021.png", "metadata/rollout_172/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0022.png", "metadata/rollout_172/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0023.png", "metadata/rollout_172/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_172/agentview/0024.png", "metadata/rollout_172/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0025.png", "metadata/rollout_172/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0026.png", "metadata/rollout_172/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0027.png", "metadata/rollout_172/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0028.png", "metadata/rollout_172/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_172/agentview/0029.png", "metadata/rollout_172/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0030.png", "metadata/rollout_172/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0031.png", "metadata/rollout_172/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0032.png", "metadata/rollout_172/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0033.png", "metadata/rollout_172/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0034.png", "metadata/rollout_172/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0035.png", "metadata/rollout_172/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_172/agentview/0036.png", "metadata/rollout_172/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0037.png", "metadata/rollout_172/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0038.png", "metadata/rollout_172/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0039.png", "metadata/rollout_172/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_172/agentview/0040.png", "metadata/rollout_172/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0041.png", "metadata/rollout_172/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0042.png", "metadata/rollout_172/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0043.png", "metadata/rollout_172/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0044.png", "metadata/rollout_172/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0045.png", "metadata/rollout_172/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0046.png", "metadata/rollout_172/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0047.png", "metadata/rollout_172/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0048.png", "metadata/rollout_172/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0049.png", "metadata/rollout_172/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_172/agentview/0050.png", "metadata/rollout_172/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0051.png", "metadata/rollout_172/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0052.png", "metadata/rollout_172/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0053.png", "metadata/rollout_172/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_172/agentview/0054.png", "metadata/rollout_172/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_172/agentview/0055.png", "metadata/rollout_172/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0056.png", "metadata/rollout_172/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0057.png", "metadata/rollout_172/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0058.png", "metadata/rollout_172/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_172/agentview/0059.png", "metadata/rollout_172/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_172/agentview/0059.png", "metadata/rollout_172/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_173/agentview/0000.png", "metadata/rollout_173/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_173/agentview/0001.png", "metadata/rollout_173/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_173/agentview/0002.png", "metadata/rollout_173/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_173/agentview/0003.png", "metadata/rollout_173/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0004.png", "metadata/rollout_173/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0005.png", "metadata/rollout_173/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0006.png", "metadata/rollout_173/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0007.png", "metadata/rollout_173/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_173/agentview/0008.png", "metadata/rollout_173/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_173/agentview/0009.png", "metadata/rollout_173/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_173/agentview/0010.png", "metadata/rollout_173/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0011.png", "metadata/rollout_173/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0012.png", "metadata/rollout_173/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0013.png", "metadata/rollout_173/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0014.png", "metadata/rollout_173/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0015.png", "metadata/rollout_173/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0016.png", "metadata/rollout_173/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0017.png", "metadata/rollout_173/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0018.png", "metadata/rollout_173/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_173/agentview/0019.png", "metadata/rollout_173/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0020.png", "metadata/rollout_173/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0021.png", "metadata/rollout_173/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0022.png", "metadata/rollout_173/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0023.png", "metadata/rollout_173/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0024.png", "metadata/rollout_173/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0025.png", "metadata/rollout_173/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_173/agentview/0026.png", "metadata/rollout_173/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_173/agentview/0027.png", "metadata/rollout_173/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_173/agentview/0028.png", "metadata/rollout_173/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0029.png", "metadata/rollout_173/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_173/agentview/0030.png", "metadata/rollout_173/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_173/agentview/0031.png", "metadata/rollout_173/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_173/agentview/0032.png", "metadata/rollout_173/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_173/agentview/0033.png", "metadata/rollout_173/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_173/agentview/0034.png", "metadata/rollout_173/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_173/agentview/0035.png", "metadata/rollout_173/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_173/agentview/0036.png", "metadata/rollout_173/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0037.png", "metadata/rollout_173/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0038.png", "metadata/rollout_173/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0039.png", "metadata/rollout_173/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_173/agentview/0040.png", "metadata/rollout_173/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_173/agentview/0041.png", "metadata/rollout_173/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0042.png", "metadata/rollout_173/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0043.png", "metadata/rollout_173/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0044.png", "metadata/rollout_173/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_173/agentview/0045.png", "metadata/rollout_173/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_173/agentview/0045.png", "metadata/rollout_173/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0000.png", "metadata/rollout_174/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0001.png", "metadata/rollout_174/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0002.png", "metadata/rollout_174/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0003.png", "metadata/rollout_174/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0004.png", "metadata/rollout_174/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0005.png", "metadata/rollout_174/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0006.png", "metadata/rollout_174/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0007.png", "metadata/rollout_174/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0008.png", "metadata/rollout_174/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0009.png", "metadata/rollout_174/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0010.png", "metadata/rollout_174/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0011.png", "metadata/rollout_174/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0012.png", "metadata/rollout_174/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0013.png", "metadata/rollout_174/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_174/agentview/0014.png", "metadata/rollout_174/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_174/agentview/0015.png", "metadata/rollout_174/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_174/agentview/0016.png", "metadata/rollout_174/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_174/agentview/0017.png", "metadata/rollout_174/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_174/agentview/0018.png", "metadata/rollout_174/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_174/agentview/0019.png", "metadata/rollout_174/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0020.png", "metadata/rollout_174/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0021.png", "metadata/rollout_174/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0022.png", "metadata/rollout_174/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0023.png", "metadata/rollout_174/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0024.png", "metadata/rollout_174/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0025.png", "metadata/rollout_174/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0026.png", "metadata/rollout_174/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0027.png", "metadata/rollout_174/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0028.png", "metadata/rollout_174/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0029.png", "metadata/rollout_174/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_174/agentview/0030.png", "metadata/rollout_174/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0031.png", "metadata/rollout_174/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0032.png", "metadata/rollout_174/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0033.png", "metadata/rollout_174/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0034.png", "metadata/rollout_174/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0035.png", "metadata/rollout_174/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0036.png", "metadata/rollout_174/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0037.png", "metadata/rollout_174/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0038.png", "metadata/rollout_174/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0039.png", "metadata/rollout_174/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0040.png", "metadata/rollout_174/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0041.png", "metadata/rollout_174/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0042.png", "metadata/rollout_174/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0043.png", "metadata/rollout_174/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0044.png", "metadata/rollout_174/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0045.png", "metadata/rollout_174/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0046.png", "metadata/rollout_174/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0047.png", "metadata/rollout_174/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0048.png", "metadata/rollout_174/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0049.png", "metadata/rollout_174/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0050.png", "metadata/rollout_174/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0051.png", "metadata/rollout_174/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0052.png", "metadata/rollout_174/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0053.png", "metadata/rollout_174/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0054.png", "metadata/rollout_174/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0055.png", "metadata/rollout_174/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_174/agentview/0056.png", "metadata/rollout_174/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_174/agentview/0057.png", "metadata/rollout_174/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_174/agentview/0058.png", "metadata/rollout_174/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_174/agentview/0059.png", "metadata/rollout_174/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0060.png", "metadata/rollout_174/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0061.png", "metadata/rollout_174/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0062.png", "metadata/rollout_174/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_174/agentview/0063.png", "metadata/rollout_174/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_174/agentview/0063.png", "metadata/rollout_174/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_175/agentview/0000.png", "metadata/rollout_175/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_175/agentview/0001.png", "metadata/rollout_175/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_175/agentview/0002.png", "metadata/rollout_175/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0003.png", "metadata/rollout_175/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0004.png", "metadata/rollout_175/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0005.png", "metadata/rollout_175/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0006.png", "metadata/rollout_175/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0007.png", "metadata/rollout_175/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0008.png", "metadata/rollout_175/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0009.png", "metadata/rollout_175/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0010.png", "metadata/rollout_175/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0011.png", "metadata/rollout_175/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0012.png", "metadata/rollout_175/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_175/agentview/0013.png", "metadata/rollout_175/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_175/agentview/0014.png", "metadata/rollout_175/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_175/agentview/0015.png", "metadata/rollout_175/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_175/agentview/0016.png", "metadata/rollout_175/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0017.png", "metadata/rollout_175/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0018.png", "metadata/rollout_175/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0019.png", "metadata/rollout_175/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0020.png", "metadata/rollout_175/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0021.png", "metadata/rollout_175/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_175/agentview/0022.png", "metadata/rollout_175/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_175/agentview/0023.png", "metadata/rollout_175/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_175/agentview/0024.png", "metadata/rollout_175/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_175/agentview/0025.png", "metadata/rollout_175/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_175/agentview/0026.png", "metadata/rollout_175/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_175/agentview/0027.png", "metadata/rollout_175/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0028.png", "metadata/rollout_175/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0029.png", "metadata/rollout_175/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0030.png", "metadata/rollout_175/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0031.png", "metadata/rollout_175/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0032.png", "metadata/rollout_175/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0033.png", "metadata/rollout_175/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0034.png", "metadata/rollout_175/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0035.png", "metadata/rollout_175/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0036.png", "metadata/rollout_175/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0037.png", "metadata/rollout_175/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0038.png", "metadata/rollout_175/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0039.png", "metadata/rollout_175/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_175/agentview/0040.png", "metadata/rollout_175/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_175/agentview/0041.png", "metadata/rollout_175/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_175/agentview/0042.png", "metadata/rollout_175/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_175/agentview/0043.png", "metadata/rollout_175/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_175/agentview/0044.png", "metadata/rollout_175/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_175/agentview/0045.png", "metadata/rollout_175/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_175/agentview/0046.png", "metadata/rollout_175/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_175/agentview/0046.png", "metadata/rollout_175/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_176/agentview/0000.png", "metadata/rollout_176/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_176/agentview/0001.png", "metadata/rollout_176/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_176/agentview/0002.png", "metadata/rollout_176/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_176/agentview/0003.png", "metadata/rollout_176/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_176/agentview/0004.png", "metadata/rollout_176/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0005.png", "metadata/rollout_176/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0006.png", "metadata/rollout_176/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0007.png", "metadata/rollout_176/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0008.png", "metadata/rollout_176/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0009.png", "metadata/rollout_176/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0010.png", "metadata/rollout_176/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0011.png", "metadata/rollout_176/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0012.png", "metadata/rollout_176/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0013.png", "metadata/rollout_176/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0014.png", "metadata/rollout_176/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0015.png", "metadata/rollout_176/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0016.png", "metadata/rollout_176/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_176/agentview/0017.png", "metadata/rollout_176/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_176/agentview/0018.png", "metadata/rollout_176/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0019.png", "metadata/rollout_176/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0020.png", "metadata/rollout_176/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0021.png", "metadata/rollout_176/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0022.png", "metadata/rollout_176/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0023.png", "metadata/rollout_176/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_176/agentview/0024.png", "metadata/rollout_176/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0025.png", "metadata/rollout_176/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0026.png", "metadata/rollout_176/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0027.png", "metadata/rollout_176/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0028.png", "metadata/rollout_176/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0029.png", "metadata/rollout_176/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0030.png", "metadata/rollout_176/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0031.png", "metadata/rollout_176/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_176/agentview/0032.png", "metadata/rollout_176/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0033.png", "metadata/rollout_176/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0034.png", "metadata/rollout_176/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0035.png", "metadata/rollout_176/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0036.png", "metadata/rollout_176/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0037.png", "metadata/rollout_176/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_176/agentview/0038.png", "metadata/rollout_176/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_176/agentview/0039.png", "metadata/rollout_176/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_176/agentview/0040.png", "metadata/rollout_176/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_176/agentview/0041.png", "metadata/rollout_176/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_176/agentview/0042.png", "metadata/rollout_176/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_176/agentview/0043.png", "metadata/rollout_176/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_176/agentview/0044.png", "metadata/rollout_176/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_176/agentview/0045.png", "metadata/rollout_176/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0046.png", "metadata/rollout_176/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0047.png", "metadata/rollout_176/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0048.png", "metadata/rollout_176/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_176/agentview/0049.png", "metadata/rollout_176/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_176/agentview/0050.png", "metadata/rollout_176/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_176/agentview/0051.png", "metadata/rollout_176/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0052.png", "metadata/rollout_176/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0053.png", "metadata/rollout_176/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0054.png", "metadata/rollout_176/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_176/agentview/0055.png", "metadata/rollout_176/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_176/agentview/0055.png", "metadata/rollout_176/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_177/agentview/0000.png", "metadata/rollout_177/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_177/agentview/0001.png", "metadata/rollout_177/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_177/agentview/0002.png", "metadata/rollout_177/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_177/agentview/0003.png", "metadata/rollout_177/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_177/agentview/0004.png", "metadata/rollout_177/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_177/agentview/0005.png", "metadata/rollout_177/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_177/agentview/0006.png", "metadata/rollout_177/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_177/agentview/0007.png", "metadata/rollout_177/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_177/agentview/0008.png", "metadata/rollout_177/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_177/agentview/0009.png", "metadata/rollout_177/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_177/agentview/0010.png", "metadata/rollout_177/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_177/agentview/0011.png", "metadata/rollout_177/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0012.png", "metadata/rollout_177/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0013.png", "metadata/rollout_177/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0014.png", "metadata/rollout_177/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0015.png", "metadata/rollout_177/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0016.png", "metadata/rollout_177/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0017.png", "metadata/rollout_177/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_177/agentview/0018.png", "metadata/rollout_177/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_177/agentview/0019.png", "metadata/rollout_177/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_177/agentview/0020.png", "metadata/rollout_177/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_177/agentview/0021.png", "metadata/rollout_177/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0022.png", "metadata/rollout_177/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0023.png", "metadata/rollout_177/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0024.png", "metadata/rollout_177/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0025.png", "metadata/rollout_177/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0026.png", "metadata/rollout_177/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0027.png", "metadata/rollout_177/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_177/agentview/0028.png", "metadata/rollout_177/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0029.png", "metadata/rollout_177/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0030.png", "metadata/rollout_177/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0031.png", "metadata/rollout_177/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0032.png", "metadata/rollout_177/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0033.png", "metadata/rollout_177/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0034.png", "metadata/rollout_177/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_177/agentview/0035.png", "metadata/rollout_177/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_177/agentview/0036.png", "metadata/rollout_177/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0037.png", "metadata/rollout_177/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0038.png", "metadata/rollout_177/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0039.png", "metadata/rollout_177/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0040.png", "metadata/rollout_177/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0041.png", "metadata/rollout_177/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0042.png", "metadata/rollout_177/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0043.png", "metadata/rollout_177/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0044.png", "metadata/rollout_177/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0045.png", "metadata/rollout_177/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0046.png", "metadata/rollout_177/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0047.png", "metadata/rollout_177/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0048.png", "metadata/rollout_177/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_177/agentview/0049.png", "metadata/rollout_177/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0050.png", "metadata/rollout_177/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0051.png", "metadata/rollout_177/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0052.png", "metadata/rollout_177/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_177/agentview/0053.png", "metadata/rollout_177/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_177/agentview/0054.png", "metadata/rollout_177/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0055.png", "metadata/rollout_177/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0056.png", "metadata/rollout_177/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0057.png", "metadata/rollout_177/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_177/agentview/0058.png", "metadata/rollout_177/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_177/agentview/0058.png", "metadata/rollout_177/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0000.png", "metadata/rollout_178/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0001.png", "metadata/rollout_178/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0002.png", "metadata/rollout_178/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0003.png", "metadata/rollout_178/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0004.png", "metadata/rollout_178/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0005.png", "metadata/rollout_178/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0006.png", "metadata/rollout_178/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_178/agentview/0007.png", "metadata/rollout_178/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_178/agentview/0008.png", "metadata/rollout_178/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_178/agentview/0009.png", "metadata/rollout_178/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_178/agentview/0010.png", "metadata/rollout_178/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_178/agentview/0011.png", "metadata/rollout_178/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_178/agentview/0012.png", "metadata/rollout_178/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0013.png", "metadata/rollout_178/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0014.png", "metadata/rollout_178/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0015.png", "metadata/rollout_178/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0016.png", "metadata/rollout_178/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0017.png", "metadata/rollout_178/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0018.png", "metadata/rollout_178/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0019.png", "metadata/rollout_178/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_178/agentview/0020.png", "metadata/rollout_178/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0021.png", "metadata/rollout_178/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0022.png", "metadata/rollout_178/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0023.png", "metadata/rollout_178/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0024.png", "metadata/rollout_178/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_178/agentview/0025.png", "metadata/rollout_178/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_178/agentview/0026.png", "metadata/rollout_178/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_178/agentview/0027.png", "metadata/rollout_178/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_178/agentview/0028.png", "metadata/rollout_178/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_178/agentview/0029.png", "metadata/rollout_178/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_178/agentview/0030.png", "metadata/rollout_178/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0031.png", "metadata/rollout_178/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0032.png", "metadata/rollout_178/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0033.png", "metadata/rollout_178/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0034.png", "metadata/rollout_178/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0035.png", "metadata/rollout_178/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0036.png", "metadata/rollout_178/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0037.png", "metadata/rollout_178/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_178/agentview/0038.png", "metadata/rollout_178/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0039.png", "metadata/rollout_178/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0040.png", "metadata/rollout_178/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0041.png", "metadata/rollout_178/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0042.png", "metadata/rollout_178/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0043.png", "metadata/rollout_178/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0044.png", "metadata/rollout_178/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0045.png", "metadata/rollout_178/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0046.png", "metadata/rollout_178/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0047.png", "metadata/rollout_178/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0048.png", "metadata/rollout_178/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0049.png", "metadata/rollout_178/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_178/agentview/0050.png", "metadata/rollout_178/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0051.png", "metadata/rollout_178/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0052.png", "metadata/rollout_178/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0053.png", "metadata/rollout_178/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_178/agentview/0054.png", "metadata/rollout_178/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_178/agentview/0055.png", "metadata/rollout_178/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_178/agentview/0056.png", "metadata/rollout_178/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_178/agentview/0057.png", "metadata/rollout_178/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_178/agentview/0058.png", "metadata/rollout_178/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_178/agentview/0059.png", "metadata/rollout_178/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_178/agentview/0059.png", "metadata/rollout_178/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_179/agentview/0000.png", "metadata/rollout_179/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0001.png", "metadata/rollout_179/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_179/agentview/0002.png", "metadata/rollout_179/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_179/agentview/0003.png", "metadata/rollout_179/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_179/agentview/0004.png", "metadata/rollout_179/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0005.png", "metadata/rollout_179/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0006.png", "metadata/rollout_179/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0007.png", "metadata/rollout_179/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0008.png", "metadata/rollout_179/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0009.png", "metadata/rollout_179/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0010.png", "metadata/rollout_179/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0011.png", "metadata/rollout_179/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0012.png", "metadata/rollout_179/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0013.png", "metadata/rollout_179/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0014.png", "metadata/rollout_179/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_179/agentview/0015.png", "metadata/rollout_179/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_179/agentview/0016.png", "metadata/rollout_179/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_179/agentview/0017.png", "metadata/rollout_179/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_179/agentview/0018.png", "metadata/rollout_179/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_179/agentview/0019.png", "metadata/rollout_179/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_179/agentview/0020.png", "metadata/rollout_179/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_179/agentview/0021.png", "metadata/rollout_179/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_179/agentview/0022.png", "metadata/rollout_179/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_179/agentview/0023.png", "metadata/rollout_179/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0024.png", "metadata/rollout_179/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_179/agentview/0025.png", "metadata/rollout_179/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_179/agentview/0026.png", "metadata/rollout_179/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_179/agentview/0027.png", "metadata/rollout_179/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_179/agentview/0028.png", "metadata/rollout_179/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_179/agentview/0029.png", "metadata/rollout_179/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0030.png", "metadata/rollout_179/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0031.png", "metadata/rollout_179/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0032.png", "metadata/rollout_179/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_179/agentview/0033.png", "metadata/rollout_179/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_179/agentview/0034.png", "metadata/rollout_179/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_179/agentview/0035.png", "metadata/rollout_179/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_179/agentview/0036.png", "metadata/rollout_179/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_179/agentview/0037.png", "metadata/rollout_179/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_179/agentview/0038.png", "metadata/rollout_179/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_179/agentview/0038.png", "metadata/rollout_179/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0000.png", "metadata/rollout_180/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0001.png", "metadata/rollout_180/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0002.png", "metadata/rollout_180/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0003.png", "metadata/rollout_180/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0004.png", "metadata/rollout_180/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0005.png", "metadata/rollout_180/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0006.png", "metadata/rollout_180/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0007.png", "metadata/rollout_180/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0008.png", "metadata/rollout_180/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0009.png", "metadata/rollout_180/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_180/agentview/0010.png", "metadata/rollout_180/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_180/agentview/0011.png", "metadata/rollout_180/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_180/agentview/0012.png", "metadata/rollout_180/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_180/agentview/0013.png", "metadata/rollout_180/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_180/agentview/0014.png", "metadata/rollout_180/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_180/agentview/0015.png", "metadata/rollout_180/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_180/agentview/0016.png", "metadata/rollout_180/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_180/agentview/0017.png", "metadata/rollout_180/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0018.png", "metadata/rollout_180/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0019.png", "metadata/rollout_180/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0020.png", "metadata/rollout_180/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0021.png", "metadata/rollout_180/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0022.png", "metadata/rollout_180/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0023.png", "metadata/rollout_180/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0024.png", "metadata/rollout_180/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0025.png", "metadata/rollout_180/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0026.png", "metadata/rollout_180/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_180/agentview/0027.png", "metadata/rollout_180/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0028.png", "metadata/rollout_180/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0029.png", "metadata/rollout_180/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0030.png", "metadata/rollout_180/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_180/agentview/0031.png", "metadata/rollout_180/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_180/agentview/0032.png", "metadata/rollout_180/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_180/agentview/0033.png", "metadata/rollout_180/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_180/agentview/0034.png", "metadata/rollout_180/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_180/agentview/0035.png", "metadata/rollout_180/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_180/agentview/0036.png", "metadata/rollout_180/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_180/agentview/0037.png", "metadata/rollout_180/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_180/agentview/0038.png", "metadata/rollout_180/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_180/agentview/0039.png", "metadata/rollout_180/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_180/agentview/0040.png", "metadata/rollout_180/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0041.png", "metadata/rollout_180/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_180/agentview/0042.png", "metadata/rollout_180/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_180/agentview/0043.png", "metadata/rollout_180/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_180/agentview/0044.png", "metadata/rollout_180/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_180/agentview/0045.png", "metadata/rollout_180/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_180/agentview/0046.png", "metadata/rollout_180/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_180/agentview/0047.png", "metadata/rollout_180/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_180/agentview/0048.png", "metadata/rollout_180/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0049.png", "metadata/rollout_180/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0050.png", "metadata/rollout_180/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0051.png", "metadata/rollout_180/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_180/agentview/0052.png", "metadata/rollout_180/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_180/agentview/0053.png", "metadata/rollout_180/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_180/agentview/0054.png", "metadata/rollout_180/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_180/agentview/0055.png", "metadata/rollout_180/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_180/agentview/0056.png", "metadata/rollout_180/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_180/agentview/0057.png", "metadata/rollout_180/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_RIGHT, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_180/agentview/0058.png", "metadata/rollout_180/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_180/agentview/0059.png", "metadata/rollout_180/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_180/agentview/0060.png", "metadata/rollout_180/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_180/agentview/0061.png", "metadata/rollout_180/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_180/agentview/0061.png", "metadata/rollout_180/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0000.png", "metadata/rollout_181/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0001.png", "metadata/rollout_181/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0002.png", "metadata/rollout_181/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0003.png", "metadata/rollout_181/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_181/agentview/0004.png", "metadata/rollout_181/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_181/agentview/0005.png", "metadata/rollout_181/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_181/agentview/0006.png", "metadata/rollout_181/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_181/agentview/0007.png", "metadata/rollout_181/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0008.png", "metadata/rollout_181/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0009.png", "metadata/rollout_181/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0010.png", "metadata/rollout_181/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0011.png", "metadata/rollout_181/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0012.png", "metadata/rollout_181/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0013.png", "metadata/rollout_181/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0014.png", "metadata/rollout_181/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0015.png", "metadata/rollout_181/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0016.png", "metadata/rollout_181/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0017.png", "metadata/rollout_181/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0018.png", "metadata/rollout_181/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0019.png", "metadata/rollout_181/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0020.png", "metadata/rollout_181/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0021.png", "metadata/rollout_181/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0022.png", "metadata/rollout_181/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0023.png", "metadata/rollout_181/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0024.png", "metadata/rollout_181/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_181/agentview/0025.png", "metadata/rollout_181/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0026.png", "metadata/rollout_181/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0027.png", "metadata/rollout_181/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0028.png", "metadata/rollout_181/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0029.png", "metadata/rollout_181/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_181/agentview/0030.png", "metadata/rollout_181/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_181/agentview/0031.png", "metadata/rollout_181/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0032.png", "metadata/rollout_181/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0033.png", "metadata/rollout_181/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0034.png", "metadata/rollout_181/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0035.png", "metadata/rollout_181/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0036.png", "metadata/rollout_181/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0037.png", "metadata/rollout_181/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0038.png", "metadata/rollout_181/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0039.png", "metadata/rollout_181/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_181/agentview/0040.png", "metadata/rollout_181/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0041.png", "metadata/rollout_181/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_181/agentview/0042.png", "metadata/rollout_181/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_181/agentview/0043.png", "metadata/rollout_181/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_181/agentview/0044.png", "metadata/rollout_181/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0045.png", "metadata/rollout_181/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0046.png", "metadata/rollout_181/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0047.png", "metadata/rollout_181/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0048.png", "metadata/rollout_181/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0049.png", "metadata/rollout_181/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0050.png", "metadata/rollout_181/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0051.png", "metadata/rollout_181/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0052.png", "metadata/rollout_181/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0053.png", "metadata/rollout_181/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0054.png", "metadata/rollout_181/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0055.png", "metadata/rollout_181/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0056.png", "metadata/rollout_181/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0057.png", "metadata/rollout_181/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0058.png", "metadata/rollout_181/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0059.png", "metadata/rollout_181/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0060.png", "metadata/rollout_181/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0061.png", "metadata/rollout_181/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_181/agentview/0062.png", "metadata/rollout_181/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0063.png", "metadata/rollout_181/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0064.png", "metadata/rollout_181/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0065.png", "metadata/rollout_181/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_181/agentview/0066.png", "metadata/rollout_181/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_181/agentview/0067.png", "metadata/rollout_181/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0068.png", "metadata/rollout_181/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0069.png", "metadata/rollout_181/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0070.png", "metadata/rollout_181/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_181/agentview/0071.png", "metadata/rollout_181/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_181/agentview/0071.png", "metadata/rollout_181/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0000.png", "metadata/rollout_182/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0001.png", "metadata/rollout_182/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0002.png", "metadata/rollout_182/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0003.png", "metadata/rollout_182/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0004.png", "metadata/rollout_182/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0005.png", "metadata/rollout_182/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0006.png", "metadata/rollout_182/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_182/agentview/0007.png", "metadata/rollout_182/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_182/agentview/0008.png", "metadata/rollout_182/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_182/agentview/0009.png", "metadata/rollout_182/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_182/agentview/0010.png", "metadata/rollout_182/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_182/agentview/0011.png", "metadata/rollout_182/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_182/agentview/0012.png", "metadata/rollout_182/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0013.png", "metadata/rollout_182/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0014.png", "metadata/rollout_182/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0015.png", "metadata/rollout_182/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0016.png", "metadata/rollout_182/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0017.png", "metadata/rollout_182/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0018.png", "metadata/rollout_182/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0019.png", "metadata/rollout_182/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0020.png", "metadata/rollout_182/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0021.png", "metadata/rollout_182/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_182/agentview/0022.png", "metadata/rollout_182/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0023.png", "metadata/rollout_182/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_182/agentview/0024.png", "metadata/rollout_182/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0025.png", "metadata/rollout_182/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0026.png", "metadata/rollout_182/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0027.png", "metadata/rollout_182/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_182/agentview/0028.png", "metadata/rollout_182/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_182/agentview/0029.png", "metadata/rollout_182/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_182/agentview/0030.png", "metadata/rollout_182/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_182/agentview/0031.png", "metadata/rollout_182/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_182/agentview/0032.png", "metadata/rollout_182/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_182/agentview/0033.png", "metadata/rollout_182/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_182/agentview/0034.png", "metadata/rollout_182/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0035.png", "metadata/rollout_182/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0036.png", "metadata/rollout_182/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0037.png", "metadata/rollout_182/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_182/agentview/0038.png", "metadata/rollout_182/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0039.png", "metadata/rollout_182/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0040.png", "metadata/rollout_182/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0041.png", "metadata/rollout_182/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0042.png", "metadata/rollout_182/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0043.png", "metadata/rollout_182/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0044.png", "metadata/rollout_182/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0045.png", "metadata/rollout_182/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0046.png", "metadata/rollout_182/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0047.png", "metadata/rollout_182/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0048.png", "metadata/rollout_182/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_182/agentview/0049.png", "metadata/rollout_182/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0050.png", "metadata/rollout_182/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0051.png", "metadata/rollout_182/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0052.png", "metadata/rollout_182/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_182/agentview/0053.png", "metadata/rollout_182/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_182/agentview/0054.png", "metadata/rollout_182/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_182/agentview/0055.png", "metadata/rollout_182/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_182/agentview/0056.png", "metadata/rollout_182/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_182/agentview/0057.png", "metadata/rollout_182/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_182/agentview/0057.png", "metadata/rollout_182/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0000.png", "metadata/rollout_183/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0001.png", "metadata/rollout_183/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0002.png", "metadata/rollout_183/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0003.png", "metadata/rollout_183/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0004.png", "metadata/rollout_183/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0005.png", "metadata/rollout_183/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0006.png", "metadata/rollout_183/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0007.png", "metadata/rollout_183/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0008.png", "metadata/rollout_183/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0009.png", "metadata/rollout_183/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0010.png", "metadata/rollout_183/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0011.png", "metadata/rollout_183/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0012.png", "metadata/rollout_183/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0013.png", "metadata/rollout_183/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0014.png", "metadata/rollout_183/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0015.png", "metadata/rollout_183/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0016.png", "metadata/rollout_183/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0017.png", "metadata/rollout_183/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0018.png", "metadata/rollout_183/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0019.png", "metadata/rollout_183/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_183/agentview/0020.png", "metadata/rollout_183/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0021.png", "metadata/rollout_183/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0022.png", "metadata/rollout_183/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0023.png", "metadata/rollout_183/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0024.png", "metadata/rollout_183/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0025.png", "metadata/rollout_183/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0026.png", "metadata/rollout_183/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0027.png", "metadata/rollout_183/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0028.png", "metadata/rollout_183/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0029.png", "metadata/rollout_183/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0030.png", "metadata/rollout_183/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0031.png", "metadata/rollout_183/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0032.png", "metadata/rollout_183/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_183/agentview/0033.png", "metadata/rollout_183/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0034.png", "metadata/rollout_183/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0035.png", "metadata/rollout_183/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0036.png", "metadata/rollout_183/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0037.png", "metadata/rollout_183/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_183/agentview/0038.png", "metadata/rollout_183/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0039.png", "metadata/rollout_183/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0040.png", "metadata/rollout_183/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0041.png", "metadata/rollout_183/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0042.png", "metadata/rollout_183/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0043.png", "metadata/rollout_183/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0044.png", "metadata/rollout_183/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_183/agentview/0045.png", "metadata/rollout_183/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0046.png", "metadata/rollout_183/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0047.png", "metadata/rollout_183/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_183/agentview/0048.png", "metadata/rollout_183/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_RIGHT, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_183/agentview/0049.png", "metadata/rollout_183/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0050.png", "metadata/rollout_183/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0051.png", "metadata/rollout_183/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0052.png", "metadata/rollout_183/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0053.png", "metadata/rollout_183/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0054.png", "metadata/rollout_183/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0055.png", "metadata/rollout_183/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0056.png", "metadata/rollout_183/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0057.png", "metadata/rollout_183/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0058.png", "metadata/rollout_183/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0059.png", "metadata/rollout_183/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_183/agentview/0060.png", "metadata/rollout_183/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_183/agentview/0061.png", "metadata/rollout_183/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0062.png", "metadata/rollout_183/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0063.png", "metadata/rollout_183/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0064.png", "metadata/rollout_183/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_183/agentview/0065.png", "metadata/rollout_183/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_183/agentview/0066.png", "metadata/rollout_183/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0067.png", "metadata/rollout_183/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0068.png", "metadata/rollout_183/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0069.png", "metadata/rollout_183/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_183/agentview/0070.png", "metadata/rollout_183/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_183/agentview/0070.png", "metadata/rollout_183/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0000.png", "metadata/rollout_184/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0001.png", "metadata/rollout_184/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0002.png", "metadata/rollout_184/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0003.png", "metadata/rollout_184/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0004.png", "metadata/rollout_184/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0005.png", "metadata/rollout_184/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0006.png", "metadata/rollout_184/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0007.png", "metadata/rollout_184/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_184/agentview/0008.png", "metadata/rollout_184/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_184/agentview/0009.png", "metadata/rollout_184/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_184/agentview/0010.png", "metadata/rollout_184/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0011.png", "metadata/rollout_184/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0012.png", "metadata/rollout_184/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0013.png", "metadata/rollout_184/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0014.png", "metadata/rollout_184/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0015.png", "metadata/rollout_184/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0016.png", "metadata/rollout_184/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0017.png", "metadata/rollout_184/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0018.png", "metadata/rollout_184/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0019.png", "metadata/rollout_184/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0020.png", "metadata/rollout_184/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0021.png", "metadata/rollout_184/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0022.png", "metadata/rollout_184/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_184/agentview/0023.png", "metadata/rollout_184/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0024.png", "metadata/rollout_184/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0025.png", "metadata/rollout_184/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0026.png", "metadata/rollout_184/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0027.png", "metadata/rollout_184/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_184/agentview/0028.png", "metadata/rollout_184/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0029.png", "metadata/rollout_184/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0030.png", "metadata/rollout_184/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0031.png", "metadata/rollout_184/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0032.png", "metadata/rollout_184/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0033.png", "metadata/rollout_184/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0034.png", "metadata/rollout_184/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_184/agentview/0035.png", "metadata/rollout_184/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0036.png", "metadata/rollout_184/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_184/agentview/0037.png", "metadata/rollout_184/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_184/agentview/0038.png", "metadata/rollout_184/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_DOWN, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_184/agentview/0039.png", "metadata/rollout_184/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_RIGHT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_184/agentview/0040.png", "metadata/rollout_184/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_184/agentview/0041.png", "metadata/rollout_184/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_184/agentview/0042.png", "metadata/rollout_184/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_184/agentview/0043.png", "metadata/rollout_184/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0044.png", "metadata/rollout_184/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0045.png", "metadata/rollout_184/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0046.png", "metadata/rollout_184/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_184/agentview/0047.png", "metadata/rollout_184/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_184/agentview/0048.png", "metadata/rollout_184/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0049.png", "metadata/rollout_184/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0050.png", "metadata/rollout_184/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0051.png", "metadata/rollout_184/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_184/agentview/0052.png", "metadata/rollout_184/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_184/agentview/0052.png", "metadata/rollout_184/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_185/agentview/0000.png", "metadata/rollout_185/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_185/agentview/0001.png", "metadata/rollout_185/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_185/agentview/0002.png", "metadata/rollout_185/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_185/agentview/0003.png", "metadata/rollout_185/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_185/agentview/0004.png", "metadata/rollout_185/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0005.png", "metadata/rollout_185/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0006.png", "metadata/rollout_185/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0007.png", "metadata/rollout_185/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0008.png", "metadata/rollout_185/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0009.png", "metadata/rollout_185/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0010.png", "metadata/rollout_185/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0011.png", "metadata/rollout_185/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0012.png", "metadata/rollout_185/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0013.png", "metadata/rollout_185/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0014.png", "metadata/rollout_185/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0015.png", "metadata/rollout_185/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0016.png", "metadata/rollout_185/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0017.png", "metadata/rollout_185/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0018.png", "metadata/rollout_185/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_185/agentview/0019.png", "metadata/rollout_185/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0020.png", "metadata/rollout_185/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0021.png", "metadata/rollout_185/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0022.png", "metadata/rollout_185/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0023.png", "metadata/rollout_185/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_185/agentview/0024.png", "metadata/rollout_185/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0025.png", "metadata/rollout_185/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0026.png", "metadata/rollout_185/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0027.png", "metadata/rollout_185/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_185/agentview/0028.png", "metadata/rollout_185/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0029.png", "metadata/rollout_185/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0030.png", "metadata/rollout_185/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0031.png", "metadata/rollout_185/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0032.png", "metadata/rollout_185/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0033.png", "metadata/rollout_185/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0034.png", "metadata/rollout_185/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_185/agentview/0035.png", "metadata/rollout_185/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_185/agentview/0036.png", "metadata/rollout_185/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_185/agentview/0037.png", "metadata/rollout_185/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_185/agentview/0038.png", "metadata/rollout_185/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0039.png", "metadata/rollout_185/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_185/agentview/0040.png", "metadata/rollout_185/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_185/agentview/0041.png", "metadata/rollout_185/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_185/agentview/0042.png", "metadata/rollout_185/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_185/agentview/0043.png", "metadata/rollout_185/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_185/agentview/0044.png", "metadata/rollout_185/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_185/agentview/0045.png", "metadata/rollout_185/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0046.png", "metadata/rollout_185/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0047.png", "metadata/rollout_185/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0048.png", "metadata/rollout_185/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_185/agentview/0049.png", "metadata/rollout_185/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_185/agentview/0050.png", "metadata/rollout_185/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_185/agentview/0051.png", "metadata/rollout_185/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0052.png", "metadata/rollout_185/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0053.png", "metadata/rollout_185/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0054.png", "metadata/rollout_185/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_185/agentview/0055.png", "metadata/rollout_185/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_185/agentview/0056.png", "metadata/rollout_185/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_185/agentview/0056.png", "metadata/rollout_185/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_186/agentview/0000.png", "metadata/rollout_186/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0001.png", "metadata/rollout_186/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_186/agentview/0002.png", "metadata/rollout_186/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_186/agentview/0003.png", "metadata/rollout_186/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_186/agentview/0004.png", "metadata/rollout_186/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_186/agentview/0005.png", "metadata/rollout_186/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_186/agentview/0006.png", "metadata/rollout_186/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_186/agentview/0007.png", "metadata/rollout_186/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_186/agentview/0008.png", "metadata/rollout_186/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_186/agentview/0009.png", "metadata/rollout_186/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0010.png", "metadata/rollout_186/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0011.png", "metadata/rollout_186/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0012.png", "metadata/rollout_186/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0013.png", "metadata/rollout_186/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0014.png", "metadata/rollout_186/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0015.png", "metadata/rollout_186/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0016.png", "metadata/rollout_186/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0017.png", "metadata/rollout_186/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0018.png", "metadata/rollout_186/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0019.png", "metadata/rollout_186/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0020.png", "metadata/rollout_186/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0021.png", "metadata/rollout_186/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0022.png", "metadata/rollout_186/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_186/agentview/0023.png", "metadata/rollout_186/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0024.png", "metadata/rollout_186/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0025.png", "metadata/rollout_186/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0026.png", "metadata/rollout_186/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0027.png", "metadata/rollout_186/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0028.png", "metadata/rollout_186/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0029.png", "metadata/rollout_186/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_186/agentview/0030.png", "metadata/rollout_186/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_186/agentview/0031.png", "metadata/rollout_186/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0032.png", "metadata/rollout_186/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_186/agentview/0033.png", "metadata/rollout_186/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_186/agentview/0034.png", "metadata/rollout_186/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_186/agentview/0035.png", "metadata/rollout_186/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_186/agentview/0036.png", "metadata/rollout_186/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_186/agentview/0037.png", "metadata/rollout_186/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0038.png", "metadata/rollout_186/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0039.png", "metadata/rollout_186/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0040.png", "metadata/rollout_186/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0041.png", "metadata/rollout_186/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_186/agentview/0042.png", "metadata/rollout_186/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_186/agentview/0043.png", "metadata/rollout_186/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_186/agentview/0044.png", "metadata/rollout_186/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0045.png", "metadata/rollout_186/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0046.png", "metadata/rollout_186/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0047.png", "metadata/rollout_186/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0048.png", "metadata/rollout_186/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_186/agentview/0049.png", "metadata/rollout_186/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_186/agentview/0049.png", "metadata/rollout_186/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_187/agentview/0000.png", "metadata/rollout_187/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_187/agentview/0001.png", "metadata/rollout_187/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_187/agentview/0002.png", "metadata/rollout_187/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_187/agentview/0003.png", "metadata/rollout_187/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_187/agentview/0004.png", "metadata/rollout_187/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_187/agentview/0005.png", "metadata/rollout_187/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_187/agentview/0006.png", "metadata/rollout_187/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_187/agentview/0007.png", "metadata/rollout_187/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_187/agentview/0008.png", "metadata/rollout_187/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_187/agentview/0009.png", "metadata/rollout_187/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_187/agentview/0010.png", "metadata/rollout_187/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_187/agentview/0011.png", "metadata/rollout_187/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0012.png", "metadata/rollout_187/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0013.png", "metadata/rollout_187/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0014.png", "metadata/rollout_187/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0015.png", "metadata/rollout_187/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0016.png", "metadata/rollout_187/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0017.png", "metadata/rollout_187/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0018.png", "metadata/rollout_187/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_187/agentview/0019.png", "metadata/rollout_187/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0020.png", "metadata/rollout_187/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0021.png", "metadata/rollout_187/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0022.png", "metadata/rollout_187/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0023.png", "metadata/rollout_187/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0024.png", "metadata/rollout_187/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_187/agentview/0025.png", "metadata/rollout_187/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0026.png", "metadata/rollout_187/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0027.png", "metadata/rollout_187/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0028.png", "metadata/rollout_187/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0029.png", "metadata/rollout_187/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0030.png", "metadata/rollout_187/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0031.png", "metadata/rollout_187/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_187/agentview/0032.png", "metadata/rollout_187/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_187/agentview/0033.png", "metadata/rollout_187/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_187/agentview/0034.png", "metadata/rollout_187/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_187/agentview/0035.png", "metadata/rollout_187/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0036.png", "metadata/rollout_187/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_187/agentview/0037.png", "metadata/rollout_187/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_187/agentview/0038.png", "metadata/rollout_187/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_187/agentview/0039.png", "metadata/rollout_187/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_187/agentview/0040.png", "metadata/rollout_187/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_187/agentview/0041.png", "metadata/rollout_187/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_187/agentview/0042.png", "metadata/rollout_187/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_187/agentview/0043.png", "metadata/rollout_187/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0044.png", "metadata/rollout_187/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0045.png", "metadata/rollout_187/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0046.png", "metadata/rollout_187/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0047.png", "metadata/rollout_187/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_187/agentview/0048.png", "metadata/rollout_187/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_187/agentview/0049.png", "metadata/rollout_187/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0050.png", "metadata/rollout_187/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0051.png", "metadata/rollout_187/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0052.png", "metadata/rollout_187/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0053.png", "metadata/rollout_187/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_187/agentview/0054.png", "metadata/rollout_187/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_187/agentview/0054.png", "metadata/rollout_187/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_188/agentview/0000.png", "metadata/rollout_188/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_188/agentview/0001.png", "metadata/rollout_188/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_188/agentview/0002.png", "metadata/rollout_188/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_188/agentview/0003.png", "metadata/rollout_188/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_188/agentview/0004.png", "metadata/rollout_188/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_188/agentview/0005.png", "metadata/rollout_188/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_188/agentview/0006.png", "metadata/rollout_188/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0007.png", "metadata/rollout_188/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0008.png", "metadata/rollout_188/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0009.png", "metadata/rollout_188/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0010.png", "metadata/rollout_188/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0011.png", "metadata/rollout_188/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0012.png", "metadata/rollout_188/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_188/agentview/0013.png", "metadata/rollout_188/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_188/agentview/0014.png", "metadata/rollout_188/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_188/agentview/0015.png", "metadata/rollout_188/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_188/agentview/0016.png", "metadata/rollout_188/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_188/agentview/0017.png", "metadata/rollout_188/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_188/agentview/0018.png", "metadata/rollout_188/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0019.png", "metadata/rollout_188/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0020.png", "metadata/rollout_188/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_188/agentview/0021.png", "metadata/rollout_188/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_188/agentview/0022.png", "metadata/rollout_188/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0023.png", "metadata/rollout_188/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0024.png", "metadata/rollout_188/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0025.png", "metadata/rollout_188/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0026.png", "metadata/rollout_188/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_188/agentview/0027.png", "metadata/rollout_188/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0028.png", "metadata/rollout_188/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0029.png", "metadata/rollout_188/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0030.png", "metadata/rollout_188/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0031.png", "metadata/rollout_188/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0032.png", "metadata/rollout_188/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0033.png", "metadata/rollout_188/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_188/agentview/0034.png", "metadata/rollout_188/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0035.png", "metadata/rollout_188/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_188/agentview/0036.png", "metadata/rollout_188/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_188/agentview/0037.png", "metadata/rollout_188/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_188/agentview/0038.png", "metadata/rollout_188/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_188/agentview/0039.png", "metadata/rollout_188/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_188/agentview/0040.png", "metadata/rollout_188/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_188/agentview/0041.png", "metadata/rollout_188/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0042.png", "metadata/rollout_188/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0043.png", "metadata/rollout_188/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0044.png", "metadata/rollout_188/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_188/agentview/0045.png", "metadata/rollout_188/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_188/agentview/0046.png", "metadata/rollout_188/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0047.png", "metadata/rollout_188/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0048.png", "metadata/rollout_188/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0049.png", "metadata/rollout_188/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_188/agentview/0050.png", "metadata/rollout_188/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_188/agentview/0050.png", "metadata/rollout_188/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_189/agentview/0000.png", "metadata/rollout_189/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_189/agentview/0001.png", "metadata/rollout_189/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_189/agentview/0002.png", "metadata/rollout_189/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_189/agentview/0003.png", "metadata/rollout_189/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_189/agentview/0004.png", "metadata/rollout_189/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0005.png", "metadata/rollout_189/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0006.png", "metadata/rollout_189/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0007.png", "metadata/rollout_189/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0008.png", "metadata/rollout_189/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0009.png", "metadata/rollout_189/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0010.png", "metadata/rollout_189/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0011.png", "metadata/rollout_189/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0012.png", "metadata/rollout_189/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0013.png", "metadata/rollout_189/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0014.png", "metadata/rollout_189/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0015.png", "metadata/rollout_189/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0016.png", "metadata/rollout_189/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0017.png", "metadata/rollout_189/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0018.png", "metadata/rollout_189/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0019.png", "metadata/rollout_189/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0020.png", "metadata/rollout_189/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0021.png", "metadata/rollout_189/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0022.png", "metadata/rollout_189/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0023.png", "metadata/rollout_189/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_189/agentview/0024.png", "metadata/rollout_189/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0025.png", "metadata/rollout_189/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0026.png", "metadata/rollout_189/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0027.png", "metadata/rollout_189/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_189/agentview/0028.png", "metadata/rollout_189/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0029.png", "metadata/rollout_189/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0030.png", "metadata/rollout_189/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0031.png", "metadata/rollout_189/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0032.png", "metadata/rollout_189/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0033.png", "metadata/rollout_189/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0034.png", "metadata/rollout_189/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_189/agentview/0035.png", "metadata/rollout_189/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_189/agentview/0036.png", "metadata/rollout_189/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0037.png", "metadata/rollout_189/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0038.png", "metadata/rollout_189/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0039.png", "metadata/rollout_189/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0040.png", "metadata/rollout_189/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0041.png", "metadata/rollout_189/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0042.png", "metadata/rollout_189/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0043.png", "metadata/rollout_189/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0044.png", "metadata/rollout_189/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0045.png", "metadata/rollout_189/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0046.png", "metadata/rollout_189/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0047.png", "metadata/rollout_189/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0048.png", "metadata/rollout_189/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0049.png", "metadata/rollout_189/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0050.png", "metadata/rollout_189/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0051.png", "metadata/rollout_189/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0052.png", "metadata/rollout_189/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0053.png", "metadata/rollout_189/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0054.png", "metadata/rollout_189/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_189/agentview/0055.png", "metadata/rollout_189/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0056.png", "metadata/rollout_189/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0057.png", "metadata/rollout_189/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0058.png", "metadata/rollout_189/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_189/agentview/0059.png", "metadata/rollout_189/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_189/agentview/0060.png", "metadata/rollout_189/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0061.png", "metadata/rollout_189/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0062.png", "metadata/rollout_189/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0063.png", "metadata/rollout_189/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_189/agentview/0064.png", "metadata/rollout_189/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_189/agentview/0064.png", "metadata/rollout_189/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_190/agentview/0000.png", "metadata/rollout_190/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_190/agentview/0001.png", "metadata/rollout_190/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_190/agentview/0002.png", "metadata/rollout_190/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_190/agentview/0003.png", "metadata/rollout_190/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_190/agentview/0004.png", "metadata/rollout_190/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_190/agentview/0005.png", "metadata/rollout_190/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_190/agentview/0006.png", "metadata/rollout_190/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_190/agentview/0007.png", "metadata/rollout_190/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_190/agentview/0008.png", "metadata/rollout_190/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_190/agentview/0009.png", "metadata/rollout_190/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_190/agentview/0010.png", "metadata/rollout_190/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_190/agentview/0011.png", "metadata/rollout_190/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_190/agentview/0012.png", "metadata/rollout_190/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_190/agentview/0013.png", "metadata/rollout_190/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_190/agentview/0014.png", "metadata/rollout_190/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_190/agentview/0015.png", "metadata/rollout_190/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_190/agentview/0016.png", "metadata/rollout_190/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0017.png", "metadata/rollout_190/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0018.png", "metadata/rollout_190/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0019.png", "metadata/rollout_190/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0020.png", "metadata/rollout_190/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0021.png", "metadata/rollout_190/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0022.png", "metadata/rollout_190/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0023.png", "metadata/rollout_190/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0024.png", "metadata/rollout_190/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0025.png", "metadata/rollout_190/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0026.png", "metadata/rollout_190/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0027.png", "metadata/rollout_190/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_190/agentview/0028.png", "metadata/rollout_190/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_190/agentview/0029.png", "metadata/rollout_190/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_190/agentview/0030.png", "metadata/rollout_190/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_190/agentview/0031.png", "metadata/rollout_190/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_190/agentview/0032.png", "metadata/rollout_190/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_190/agentview/0033.png", "metadata/rollout_190/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_190/agentview/0034.png", "metadata/rollout_190/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0035.png", "metadata/rollout_190/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0036.png", "metadata/rollout_190/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0037.png", "metadata/rollout_190/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0038.png", "metadata/rollout_190/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0039.png", "metadata/rollout_190/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0040.png", "metadata/rollout_190/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_190/agentview/0041.png", "metadata/rollout_190/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_190/agentview/0042.png", "metadata/rollout_190/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_190/agentview/0043.png", "metadata/rollout_190/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0044.png", "metadata/rollout_190/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0045.png", "metadata/rollout_190/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0046.png", "metadata/rollout_190/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0047.png", "metadata/rollout_190/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0048.png", "metadata/rollout_190/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0049.png", "metadata/rollout_190/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0050.png", "metadata/rollout_190/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_190/agentview/0051.png", "metadata/rollout_190/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_190/agentview/0052.png", "metadata/rollout_190/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0053.png", "metadata/rollout_190/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0054.png", "metadata/rollout_190/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0055.png", "metadata/rollout_190/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_190/agentview/0056.png", "metadata/rollout_190/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_190/agentview/0057.png", "metadata/rollout_190/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_190/agentview/0058.png", "metadata/rollout_190/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_190/agentview/0059.png", "metadata/rollout_190/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_190/agentview/0060.png", "metadata/rollout_190/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_190/agentview/0061.png", "metadata/rollout_190/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_190/agentview/0061.png", "metadata/rollout_190/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0000.png", "metadata/rollout_191/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0001.png", "metadata/rollout_191/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0002.png", "metadata/rollout_191/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_191/agentview/0003.png", "metadata/rollout_191/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_191/agentview/0004.png", "metadata/rollout_191/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_191/agentview/0005.png", "metadata/rollout_191/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0006.png", "metadata/rollout_191/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0007.png", "metadata/rollout_191/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0008.png", "metadata/rollout_191/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0009.png", "metadata/rollout_191/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0010.png", "metadata/rollout_191/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_191/agentview/0011.png", "metadata/rollout_191/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_191/agentview/0012.png", "metadata/rollout_191/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0013.png", "metadata/rollout_191/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0014.png", "metadata/rollout_191/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0015.png", "metadata/rollout_191/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0016.png", "metadata/rollout_191/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0017.png", "metadata/rollout_191/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0018.png", "metadata/rollout_191/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_191/agentview/0019.png", "metadata/rollout_191/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_191/agentview/0020.png", "metadata/rollout_191/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_191/agentview/0021.png", "metadata/rollout_191/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_191/agentview/0022.png", "metadata/rollout_191/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_191/agentview/0023.png", "metadata/rollout_191/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_191/agentview/0024.png", "metadata/rollout_191/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0025.png", "metadata/rollout_191/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0026.png", "metadata/rollout_191/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0027.png", "metadata/rollout_191/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_191/agentview/0028.png", "metadata/rollout_191/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_191/agentview/0029.png", "metadata/rollout_191/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0030.png", "metadata/rollout_191/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0031.png", "metadata/rollout_191/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0032.png", "metadata/rollout_191/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0033.png", "metadata/rollout_191/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0034.png", "metadata/rollout_191/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_191/agentview/0035.png", "metadata/rollout_191/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0036.png", "metadata/rollout_191/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0037.png", "metadata/rollout_191/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0038.png", "metadata/rollout_191/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_191/agentview/0039.png", "metadata/rollout_191/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_191/agentview/0040.png", "metadata/rollout_191/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_191/agentview/0041.png", "metadata/rollout_191/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_191/agentview/0042.png", "metadata/rollout_191/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_191/agentview/0043.png", "metadata/rollout_191/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_191/agentview/0044.png", "metadata/rollout_191/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_191/agentview/0044.png", "metadata/rollout_191/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0000.png", "metadata/rollout_192/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0001.png", "metadata/rollout_192/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0002.png", "metadata/rollout_192/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0003.png", "metadata/rollout_192/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0004.png", "metadata/rollout_192/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0005.png", "metadata/rollout_192/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_192/agentview/0006.png", "metadata/rollout_192/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_192/agentview/0007.png", "metadata/rollout_192/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_192/agentview/0008.png", "metadata/rollout_192/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_192/agentview/0009.png", "metadata/rollout_192/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_192/agentview/0010.png", "metadata/rollout_192/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0011.png", "metadata/rollout_192/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0012.png", "metadata/rollout_192/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0013.png", "metadata/rollout_192/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0014.png", "metadata/rollout_192/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0015.png", "metadata/rollout_192/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0016.png", "metadata/rollout_192/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0017.png", "metadata/rollout_192/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0018.png", "metadata/rollout_192/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_192/agentview/0019.png", "metadata/rollout_192/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_192/agentview/0020.png", "metadata/rollout_192/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_192/agentview/0021.png", "metadata/rollout_192/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0022.png", "metadata/rollout_192/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0023.png", "metadata/rollout_192/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0024.png", "metadata/rollout_192/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0025.png", "metadata/rollout_192/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_192/agentview/0026.png", "metadata/rollout_192/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0027.png", "metadata/rollout_192/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0028.png", "metadata/rollout_192/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0029.png", "metadata/rollout_192/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0030.png", "metadata/rollout_192/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0031.png", "metadata/rollout_192/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0032.png", "metadata/rollout_192/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0033.png", "metadata/rollout_192/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0034.png", "metadata/rollout_192/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0035.png", "metadata/rollout_192/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0036.png", "metadata/rollout_192/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0037.png", "metadata/rollout_192/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0038.png", "metadata/rollout_192/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0039.png", "metadata/rollout_192/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0040.png", "metadata/rollout_192/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_192/agentview/0041.png", "metadata/rollout_192/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0042.png", "metadata/rollout_192/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0043.png", "metadata/rollout_192/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0044.png", "metadata/rollout_192/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0045.png", "metadata/rollout_192/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0046.png", "metadata/rollout_192/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0047.png", "metadata/rollout_192/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0048.png", "metadata/rollout_192/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0049.png", "metadata/rollout_192/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0050.png", "metadata/rollout_192/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0051.png", "metadata/rollout_192/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0052.png", "metadata/rollout_192/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0053.png", "metadata/rollout_192/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0054.png", "metadata/rollout_192/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0055.png", "metadata/rollout_192/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_192/agentview/0056.png", "metadata/rollout_192/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0057.png", "metadata/rollout_192/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0058.png", "metadata/rollout_192/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0059.png", "metadata/rollout_192/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_192/agentview/0060.png", "metadata/rollout_192/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_192/agentview/0061.png", "metadata/rollout_192/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0062.png", "metadata/rollout_192/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0063.png", "metadata/rollout_192/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0064.png", "metadata/rollout_192/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_192/agentview/0065.png", "metadata/rollout_192/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_192/agentview/0065.png", "metadata/rollout_192/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0000.png", "metadata/rollout_193/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0001.png", "metadata/rollout_193/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0002.png", "metadata/rollout_193/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0003.png", "metadata/rollout_193/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0004.png", "metadata/rollout_193/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0005.png", "metadata/rollout_193/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0006.png", "metadata/rollout_193/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0007.png", "metadata/rollout_193/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0008.png", "metadata/rollout_193/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0009.png", "metadata/rollout_193/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0010.png", "metadata/rollout_193/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_193/agentview/0011.png", "metadata/rollout_193/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_193/agentview/0012.png", "metadata/rollout_193/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_193/agentview/0013.png", "metadata/rollout_193/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_193/agentview/0014.png", "metadata/rollout_193/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_193/agentview/0015.png", "metadata/rollout_193/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0016.png", "metadata/rollout_193/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0017.png", "metadata/rollout_193/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0018.png", "metadata/rollout_193/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0019.png", "metadata/rollout_193/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0020.png", "metadata/rollout_193/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0021.png", "metadata/rollout_193/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0022.png", "metadata/rollout_193/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0023.png", "metadata/rollout_193/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0024.png", "metadata/rollout_193/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0025.png", "metadata/rollout_193/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_193/agentview/0026.png", "metadata/rollout_193/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0027.png", "metadata/rollout_193/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0028.png", "metadata/rollout_193/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0029.png", "metadata/rollout_193/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0030.png", "metadata/rollout_193/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0031.png", "metadata/rollout_193/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0032.png", "metadata/rollout_193/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0033.png", "metadata/rollout_193/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0034.png", "metadata/rollout_193/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0035.png", "metadata/rollout_193/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0036.png", "metadata/rollout_193/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0037.png", "metadata/rollout_193/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0038.png", "metadata/rollout_193/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0039.png", "metadata/rollout_193/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0040.png", "metadata/rollout_193/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0041.png", "metadata/rollout_193/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0042.png", "metadata/rollout_193/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0043.png", "metadata/rollout_193/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0044.png", "metadata/rollout_193/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0045.png", "metadata/rollout_193/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0046.png", "metadata/rollout_193/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0047.png", "metadata/rollout_193/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0048.png", "metadata/rollout_193/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0049.png", "metadata/rollout_193/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_193/agentview/0050.png", "metadata/rollout_193/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_193/agentview/0051.png", "metadata/rollout_193/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0052.png", "metadata/rollout_193/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0053.png", "metadata/rollout_193/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0054.png", "metadata/rollout_193/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_193/agentview/0055.png", "metadata/rollout_193/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_193/agentview/0056.png", "metadata/rollout_193/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0057.png", "metadata/rollout_193/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0058.png", "metadata/rollout_193/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0059.png", "metadata/rollout_193/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_193/agentview/0060.png", "metadata/rollout_193/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_193/agentview/0060.png", "metadata/rollout_193/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0000.png", "metadata/rollout_194/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0001.png", "metadata/rollout_194/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0002.png", "metadata/rollout_194/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0003.png", "metadata/rollout_194/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_194/agentview/0004.png", "metadata/rollout_194/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_194/agentview/0005.png", "metadata/rollout_194/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0006.png", "metadata/rollout_194/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0007.png", "metadata/rollout_194/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0008.png", "metadata/rollout_194/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0009.png", "metadata/rollout_194/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0010.png", "metadata/rollout_194/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0011.png", "metadata/rollout_194/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0012.png", "metadata/rollout_194/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0013.png", "metadata/rollout_194/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0014.png", "metadata/rollout_194/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0015.png", "metadata/rollout_194/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0016.png", "metadata/rollout_194/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0017.png", "metadata/rollout_194/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0018.png", "metadata/rollout_194/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0019.png", "metadata/rollout_194/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0020.png", "metadata/rollout_194/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0021.png", "metadata/rollout_194/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0022.png", "metadata/rollout_194/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_194/agentview/0023.png", "metadata/rollout_194/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0024.png", "metadata/rollout_194/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0025.png", "metadata/rollout_194/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0026.png", "metadata/rollout_194/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0027.png", "metadata/rollout_194/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0028.png", "metadata/rollout_194/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0029.png", "metadata/rollout_194/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0030.png", "metadata/rollout_194/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0031.png", "metadata/rollout_194/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_194/agentview/0032.png", "metadata/rollout_194/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_194/agentview/0033.png", "metadata/rollout_194/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0034.png", "metadata/rollout_194/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0035.png", "metadata/rollout_194/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_194/agentview/0036.png", "metadata/rollout_194/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_194/agentview/0037.png", "metadata/rollout_194/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_194/agentview/0038.png", "metadata/rollout_194/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_194/agentview/0039.png", "metadata/rollout_194/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_194/agentview/0040.png", "metadata/rollout_194/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0041.png", "metadata/rollout_194/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0042.png", "metadata/rollout_194/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0043.png", "metadata/rollout_194/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_194/agentview/0044.png", "metadata/rollout_194/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_194/agentview/0045.png", "metadata/rollout_194/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0046.png", "metadata/rollout_194/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0047.png", "metadata/rollout_194/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0048.png", "metadata/rollout_194/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_194/agentview/0049.png", "metadata/rollout_194/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_194/agentview/0049.png", "metadata/rollout_194/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0000.png", "metadata/rollout_195/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0001.png", "metadata/rollout_195/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0002.png", "metadata/rollout_195/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0003.png", "metadata/rollout_195/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0004.png", "metadata/rollout_195/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0005.png", "metadata/rollout_195/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0006.png", "metadata/rollout_195/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0007.png", "metadata/rollout_195/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_195/agentview/0008.png", "metadata/rollout_195/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_195/agentview/0009.png", "metadata/rollout_195/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_195/agentview/0010.png", "metadata/rollout_195/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_195/agentview/0011.png", "metadata/rollout_195/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_195/agentview/0012.png", "metadata/rollout_195/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_195/agentview/0013.png", "metadata/rollout_195/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_195/agentview/0014.png", "metadata/rollout_195/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0015.png", "metadata/rollout_195/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0016.png", "metadata/rollout_195/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0017.png", "metadata/rollout_195/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0018.png", "metadata/rollout_195/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0019.png", "metadata/rollout_195/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0020.png", "metadata/rollout_195/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0021.png", "metadata/rollout_195/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0022.png", "metadata/rollout_195/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0023.png", "metadata/rollout_195/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0024.png", "metadata/rollout_195/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0025.png", "metadata/rollout_195/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0026.png", "metadata/rollout_195/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0027.png", "metadata/rollout_195/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_195/agentview/0028.png", "metadata/rollout_195/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_195/agentview/0029.png", "metadata/rollout_195/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0030.png", "metadata/rollout_195/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0031.png", "metadata/rollout_195/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0032.png", "metadata/rollout_195/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0033.png", "metadata/rollout_195/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_195/agentview/0034.png", "metadata/rollout_195/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0035.png", "metadata/rollout_195/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0036.png", "metadata/rollout_195/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0037.png", "metadata/rollout_195/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0038.png", "metadata/rollout_195/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0039.png", "metadata/rollout_195/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0040.png", "metadata/rollout_195/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_195/agentview/0041.png", "metadata/rollout_195/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_195/agentview/0042.png", "metadata/rollout_195/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_195/agentview/0043.png", "metadata/rollout_195/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_195/agentview/0044.png", "metadata/rollout_195/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0045.png", "metadata/rollout_195/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_195/agentview/0046.png", "metadata/rollout_195/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_195/agentview/0047.png", "metadata/rollout_195/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_195/agentview/0048.png", "metadata/rollout_195/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_195/agentview/0049.png", "metadata/rollout_195/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_195/agentview/0050.png", "metadata/rollout_195/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_195/agentview/0051.png", "metadata/rollout_195/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_195/agentview/0052.png", "metadata/rollout_195/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_195/agentview/0053.png", "metadata/rollout_195/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0054.png", "metadata/rollout_195/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0055.png", "metadata/rollout_195/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0056.png", "metadata/rollout_195/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_195/agentview/0057.png", "metadata/rollout_195/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_195/agentview/0058.png", "metadata/rollout_195/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0059.png", "metadata/rollout_195/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0060.png", "metadata/rollout_195/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0061.png", "metadata/rollout_195/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_195/agentview/0062.png", "metadata/rollout_195/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_195/agentview/0062.png", "metadata/rollout_195/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0000.png", "metadata/rollout_196/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0001.png", "metadata/rollout_196/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0002.png", "metadata/rollout_196/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0003.png", "metadata/rollout_196/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0004.png", "metadata/rollout_196/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0005.png", "metadata/rollout_196/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0006.png", "metadata/rollout_196/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0007.png", "metadata/rollout_196/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0008.png", "metadata/rollout_196/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_196/agentview/0009.png", "metadata/rollout_196/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_196/agentview/0010.png", "metadata/rollout_196/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_196/agentview/0011.png", "metadata/rollout_196/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_196/agentview/0012.png", "metadata/rollout_196/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_196/agentview/0013.png", "metadata/rollout_196/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_196/agentview/0014.png", "metadata/rollout_196/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_196/agentview/0015.png", "metadata/rollout_196/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_196/agentview/0016.png", "metadata/rollout_196/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0017.png", "metadata/rollout_196/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0018.png", "metadata/rollout_196/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0019.png", "metadata/rollout_196/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0020.png", "metadata/rollout_196/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0021.png", "metadata/rollout_196/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0022.png", "metadata/rollout_196/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0023.png", "metadata/rollout_196/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0024.png", "metadata/rollout_196/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_196/agentview/0025.png", "metadata/rollout_196/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_196/agentview/0026.png", "metadata/rollout_196/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0027.png", "metadata/rollout_196/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0028.png", "metadata/rollout_196/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0029.png", "metadata/rollout_196/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_196/agentview/0030.png", "metadata/rollout_196/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_196/agentview/0031.png", "metadata/rollout_196/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_196/agentview/0032.png", "metadata/rollout_196/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_196/agentview/0033.png", "metadata/rollout_196/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_196/agentview/0034.png", "metadata/rollout_196/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_196/agentview/0035.png", "metadata/rollout_196/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0036.png", "metadata/rollout_196/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_196/agentview/0037.png", "metadata/rollout_196/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0038.png", "metadata/rollout_196/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0039.png", "metadata/rollout_196/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0040.png", "metadata/rollout_196/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_196/agentview/0041.png", "metadata/rollout_196/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_196/agentview/0042.png", "metadata/rollout_196/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0043.png", "metadata/rollout_196/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0044.png", "metadata/rollout_196/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0045.png", "metadata/rollout_196/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0046.png", "metadata/rollout_196/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0047.png", "metadata/rollout_196/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0048.png", "metadata/rollout_196/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0049.png", "metadata/rollout_196/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0050.png", "metadata/rollout_196/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0051.png", "metadata/rollout_196/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0052.png", "metadata/rollout_196/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_196/agentview/0053.png", "metadata/rollout_196/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0054.png", "metadata/rollout_196/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0055.png", "metadata/rollout_196/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0056.png", "metadata/rollout_196/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_196/agentview/0057.png", "metadata/rollout_196/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_196/agentview/0058.png", "metadata/rollout_196/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_196/agentview/0059.png", "metadata/rollout_196/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_196/agentview/0060.png", "metadata/rollout_196/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_196/agentview/0061.png", "metadata/rollout_196/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_196/agentview/0062.png", "metadata/rollout_196/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_196/agentview/0063.png", "metadata/rollout_196/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_196/agentview/0063.png", "metadata/rollout_196/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_197/agentview/0000.png", "metadata/rollout_197/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_197/agentview/0001.png", "metadata/rollout_197/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_197/agentview/0002.png", "metadata/rollout_197/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_197/agentview/0003.png", "metadata/rollout_197/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0004.png", "metadata/rollout_197/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0005.png", "metadata/rollout_197/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0006.png", "metadata/rollout_197/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0007.png", "metadata/rollout_197/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_197/agentview/0008.png", "metadata/rollout_197/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_197/agentview/0009.png", "metadata/rollout_197/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_197/agentview/0010.png", "metadata/rollout_197/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_197/agentview/0011.png", "metadata/rollout_197/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_197/agentview/0012.png", "metadata/rollout_197/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_197/agentview/0013.png", "metadata/rollout_197/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_197/agentview/0014.png", "metadata/rollout_197/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_197/agentview/0015.png", "metadata/rollout_197/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_197/agentview/0016.png", "metadata/rollout_197/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0017.png", "metadata/rollout_197/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0018.png", "metadata/rollout_197/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0019.png", "metadata/rollout_197/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0020.png", "metadata/rollout_197/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0021.png", "metadata/rollout_197/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0022.png", "metadata/rollout_197/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0023.png", "metadata/rollout_197/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_197/agentview/0024.png", "metadata/rollout_197/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_197/agentview/0025.png", "metadata/rollout_197/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_197/agentview/0026.png", "metadata/rollout_197/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_197/agentview/0027.png", "metadata/rollout_197/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_197/agentview/0028.png", "metadata/rollout_197/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_197/agentview/0029.png", "metadata/rollout_197/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_197/agentview/0030.png", "metadata/rollout_197/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0031.png", "metadata/rollout_197/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0032.png", "metadata/rollout_197/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_197/agentview/0033.png", "metadata/rollout_197/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0034.png", "metadata/rollout_197/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0035.png", "metadata/rollout_197/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0036.png", "metadata/rollout_197/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0037.png", "metadata/rollout_197/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0038.png", "metadata/rollout_197/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0039.png", "metadata/rollout_197/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0040.png", "metadata/rollout_197/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0041.png", "metadata/rollout_197/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0042.png", "metadata/rollout_197/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0043.png", "metadata/rollout_197/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0044.png", "metadata/rollout_197/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0045.png", "metadata/rollout_197/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0046.png", "metadata/rollout_197/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0047.png", "metadata/rollout_197/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0048.png", "metadata/rollout_197/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0049.png", "metadata/rollout_197/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_197/agentview/0050.png", "metadata/rollout_197/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_197/agentview/0051.png", "metadata/rollout_197/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0052.png", "metadata/rollout_197/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0053.png", "metadata/rollout_197/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0054.png", "metadata/rollout_197/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_197/agentview/0055.png", "metadata/rollout_197/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_197/agentview/0056.png", "metadata/rollout_197/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_197/agentview/0057.png", "metadata/rollout_197/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_197/agentview/0058.png", "metadata/rollout_197/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_197/agentview/0059.png", "metadata/rollout_197/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_197/agentview/0060.png", "metadata/rollout_197/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_197/agentview/0061.png", "metadata/rollout_197/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_197/agentview/0061.png", "metadata/rollout_197/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0000.png", "metadata/rollout_198/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0001.png", "metadata/rollout_198/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0002.png", "metadata/rollout_198/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0003.png", "metadata/rollout_198/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0004.png", "metadata/rollout_198/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0005.png", "metadata/rollout_198/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0006.png", "metadata/rollout_198/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0007.png", "metadata/rollout_198/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0008.png", "metadata/rollout_198/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0009.png", "metadata/rollout_198/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0010.png", "metadata/rollout_198/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0011.png", "metadata/rollout_198/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0012.png", "metadata/rollout_198/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0013.png", "metadata/rollout_198/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0014.png", "metadata/rollout_198/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0015.png", "metadata/rollout_198/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0016.png", "metadata/rollout_198/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0017.png", "metadata/rollout_198/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0018.png", "metadata/rollout_198/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0019.png", "metadata/rollout_198/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0020.png", "metadata/rollout_198/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0021.png", "metadata/rollout_198/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0022.png", "metadata/rollout_198/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0023.png", "metadata/rollout_198/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0024.png", "metadata/rollout_198/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_198/agentview/0025.png", "metadata/rollout_198/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0026.png", "metadata/rollout_198/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0027.png", "metadata/rollout_198/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0028.png", "metadata/rollout_198/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0029.png", "metadata/rollout_198/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_198/agentview/0030.png", "metadata/rollout_198/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0031.png", "metadata/rollout_198/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0032.png", "metadata/rollout_198/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0033.png", "metadata/rollout_198/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0034.png", "metadata/rollout_198/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0035.png", "metadata/rollout_198/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0036.png", "metadata/rollout_198/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0037.png", "metadata/rollout_198/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0038.png", "metadata/rollout_198/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0039.png", "metadata/rollout_198/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0040.png", "metadata/rollout_198/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0041.png", "metadata/rollout_198/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_198/agentview/0042.png", "metadata/rollout_198/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0043.png", "metadata/rollout_198/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0044.png", "metadata/rollout_198/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0045.png", "metadata/rollout_198/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0046.png", "metadata/rollout_198/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0047.png", "metadata/rollout_198/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0048.png", "metadata/rollout_198/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0049.png", "metadata/rollout_198/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0050.png", "metadata/rollout_198/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0051.png", "metadata/rollout_198/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0052.png", "metadata/rollout_198/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0053.png", "metadata/rollout_198/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0054.png", "metadata/rollout_198/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0055.png", "metadata/rollout_198/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0056.png", "metadata/rollout_198/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0057.png", "metadata/rollout_198/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0058.png", "metadata/rollout_198/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_198/agentview/0059.png", "metadata/rollout_198/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0060.png", "metadata/rollout_198/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0061.png", "metadata/rollout_198/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0062.png", "metadata/rollout_198/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_198/agentview/0063.png", "metadata/rollout_198/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_198/agentview/0064.png", "metadata/rollout_198/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0065.png", "metadata/rollout_198/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0066.png", "metadata/rollout_198/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0067.png", "metadata/rollout_198/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_198/agentview/0068.png", "metadata/rollout_198/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_198/agentview/0068.png", "metadata/rollout_198/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_199/agentview/0000.png", "metadata/rollout_199/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_199/agentview/0001.png", "metadata/rollout_199/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_199/agentview/0002.png", "metadata/rollout_199/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_199/agentview/0003.png", "metadata/rollout_199/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0004.png", "metadata/rollout_199/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0005.png", "metadata/rollout_199/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0006.png", "metadata/rollout_199/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0007.png", "metadata/rollout_199/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_199/agentview/0008.png", "metadata/rollout_199/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_199/agentview/0009.png", "metadata/rollout_199/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_199/agentview/0010.png", "metadata/rollout_199/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_199/agentview/0011.png", "metadata/rollout_199/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_199/agentview/0012.png", "metadata/rollout_199/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_199/agentview/0013.png", "metadata/rollout_199/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0014.png", "metadata/rollout_199/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0015.png", "metadata/rollout_199/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0016.png", "metadata/rollout_199/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0017.png", "metadata/rollout_199/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0018.png", "metadata/rollout_199/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0019.png", "metadata/rollout_199/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0020.png", "metadata/rollout_199/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0021.png", "metadata/rollout_199/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0022.png", "metadata/rollout_199/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0023.png", "metadata/rollout_199/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_199/agentview/0024.png", "metadata/rollout_199/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_199/agentview/0025.png", "metadata/rollout_199/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0026.png", "metadata/rollout_199/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0027.png", "metadata/rollout_199/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0028.png", "metadata/rollout_199/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0029.png", "metadata/rollout_199/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0030.png", "metadata/rollout_199/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0031.png", "metadata/rollout_199/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_199/agentview/0032.png", "metadata/rollout_199/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_199/agentview/0033.png", "metadata/rollout_199/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0034.png", "metadata/rollout_199/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0035.png", "metadata/rollout_199/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0036.png", "metadata/rollout_199/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0037.png", "metadata/rollout_199/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_199/agentview/0038.png", "metadata/rollout_199/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0039.png", "metadata/rollout_199/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0040.png", "metadata/rollout_199/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0041.png", "metadata/rollout_199/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0042.png", "metadata/rollout_199/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0043.png", "metadata/rollout_199/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0044.png", "metadata/rollout_199/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0045.png", "metadata/rollout_199/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0046.png", "metadata/rollout_199/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_199/agentview/0047.png", "metadata/rollout_199/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0048.png", "metadata/rollout_199/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0049.png", "metadata/rollout_199/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0050.png", "metadata/rollout_199/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_199/agentview/0051.png", "metadata/rollout_199/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_199/agentview/0052.png", "metadata/rollout_199/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0053.png", "metadata/rollout_199/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0054.png", "metadata/rollout_199/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0055.png", "metadata/rollout_199/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_199/agentview/0056.png", "metadata/rollout_199/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_199/agentview/0056.png", "metadata/rollout_199/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0000.png", "metadata/rollout_200/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0001.png", "metadata/rollout_200/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0002.png", "metadata/rollout_200/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0003.png", "metadata/rollout_200/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0004.png", "metadata/rollout_200/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0005.png", "metadata/rollout_200/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0006.png", "metadata/rollout_200/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_200/agentview/0007.png", "metadata/rollout_200/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_200/agentview/0008.png", "metadata/rollout_200/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_200/agentview/0009.png", "metadata/rollout_200/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_200/agentview/0010.png", "metadata/rollout_200/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_200/agentview/0011.png", "metadata/rollout_200/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_200/agentview/0012.png", "metadata/rollout_200/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0013.png", "metadata/rollout_200/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0014.png", "metadata/rollout_200/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0015.png", "metadata/rollout_200/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0016.png", "metadata/rollout_200/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0017.png", "metadata/rollout_200/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0018.png", "metadata/rollout_200/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0019.png", "metadata/rollout_200/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_200/agentview/0020.png", "metadata/rollout_200/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0021.png", "metadata/rollout_200/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0022.png", "metadata/rollout_200/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0023.png", "metadata/rollout_200/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0024.png", "metadata/rollout_200/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0025.png", "metadata/rollout_200/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0026.png", "metadata/rollout_200/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0027.png", "metadata/rollout_200/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0028.png", "metadata/rollout_200/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0029.png", "metadata/rollout_200/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0030.png", "metadata/rollout_200/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0031.png", "metadata/rollout_200/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_200/agentview/0032.png", "metadata/rollout_200/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0033.png", "metadata/rollout_200/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0034.png", "metadata/rollout_200/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0035.png", "metadata/rollout_200/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0036.png", "metadata/rollout_200/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0037.png", "metadata/rollout_200/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0038.png", "metadata/rollout_200/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_200/agentview/0039.png", "metadata/rollout_200/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_200/agentview/0040.png", "metadata/rollout_200/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0041.png", "metadata/rollout_200/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0042.png", "metadata/rollout_200/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0043.png", "metadata/rollout_200/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0044.png", "metadata/rollout_200/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0045.png", "metadata/rollout_200/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0046.png", "metadata/rollout_200/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_200/agentview/0047.png", "metadata/rollout_200/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_200/agentview/0048.png", "metadata/rollout_200/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_200/agentview/0049.png", "metadata/rollout_200/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_200/agentview/0050.png", "metadata/rollout_200/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0051.png", "metadata/rollout_200/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0052.png", "metadata/rollout_200/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0053.png", "metadata/rollout_200/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0054.png", "metadata/rollout_200/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0055.png", "metadata/rollout_200/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0056.png", "metadata/rollout_200/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0057.png", "metadata/rollout_200/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0058.png", "metadata/rollout_200/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0059.png", "metadata/rollout_200/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0060.png", "metadata/rollout_200/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0061.png", "metadata/rollout_200/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0062.png", "metadata/rollout_200/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0063.png", "metadata/rollout_200/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0064.png", "metadata/rollout_200/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0065.png", "metadata/rollout_200/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_200/agentview/0066.png", "metadata/rollout_200/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_200/agentview/0067.png", "metadata/rollout_200/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_200/agentview/0068.png", "metadata/rollout_200/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_200/agentview/0069.png", "metadata/rollout_200/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0070.png", "metadata/rollout_200/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0071.png", "metadata/rollout_200/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0072.png", "metadata/rollout_200/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_200/agentview/0073.png", "metadata/rollout_200/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_200/agentview/0074.png", "metadata/rollout_200/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_200/agentview/0075.png", "metadata/rollout_200/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_200/agentview/0075.png", "metadata/rollout_200/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0000.png", "metadata/rollout_201/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0001.png", "metadata/rollout_201/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0002.png", "metadata/rollout_201/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0003.png", "metadata/rollout_201/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0004.png", "metadata/rollout_201/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0005.png", "metadata/rollout_201/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0006.png", "metadata/rollout_201/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0007.png", "metadata/rollout_201/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_201/agentview/0008.png", "metadata/rollout_201/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_201/agentview/0009.png", "metadata/rollout_201/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_201/agentview/0010.png", "metadata/rollout_201/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_201/agentview/0011.png", "metadata/rollout_201/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_201/agentview/0012.png", "metadata/rollout_201/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_201/agentview/0013.png", "metadata/rollout_201/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_201/agentview/0014.png", "metadata/rollout_201/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0015.png", "metadata/rollout_201/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0016.png", "metadata/rollout_201/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0017.png", "metadata/rollout_201/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0018.png", "metadata/rollout_201/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0019.png", "metadata/rollout_201/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0020.png", "metadata/rollout_201/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0021.png", "metadata/rollout_201/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0022.png", "metadata/rollout_201/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0023.png", "metadata/rollout_201/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_201/agentview/0024.png", "metadata/rollout_201/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0025.png", "metadata/rollout_201/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0026.png", "metadata/rollout_201/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0027.png", "metadata/rollout_201/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0028.png", "metadata/rollout_201/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0029.png", "metadata/rollout_201/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0030.png", "metadata/rollout_201/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0031.png", "metadata/rollout_201/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_201/agentview/0032.png", "metadata/rollout_201/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0033.png", "metadata/rollout_201/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0034.png", "metadata/rollout_201/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0035.png", "metadata/rollout_201/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0036.png", "metadata/rollout_201/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0037.png", "metadata/rollout_201/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0038.png", "metadata/rollout_201/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_201/agentview/0039.png", "metadata/rollout_201/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0040.png", "metadata/rollout_201/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0041.png", "metadata/rollout_201/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_201/agentview/0042.png", "metadata/rollout_201/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_201/agentview/0043.png", "metadata/rollout_201/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0044.png", "metadata/rollout_201/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0045.png", "metadata/rollout_201/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0046.png", "metadata/rollout_201/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0047.png", "metadata/rollout_201/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0048.png", "metadata/rollout_201/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0049.png", "metadata/rollout_201/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0050.png", "metadata/rollout_201/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0051.png", "metadata/rollout_201/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0052.png", "metadata/rollout_201/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0053.png", "metadata/rollout_201/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0054.png", "metadata/rollout_201/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_201/agentview/0055.png", "metadata/rollout_201/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_201/agentview/0056.png", "metadata/rollout_201/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_201/agentview/0057.png", "metadata/rollout_201/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0058.png", "metadata/rollout_201/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0059.png", "metadata/rollout_201/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0060.png", "metadata/rollout_201/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_201/agentview/0061.png", "metadata/rollout_201/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_201/agentview/0062.png", "metadata/rollout_201/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_201/agentview/0062.png", "metadata/rollout_201/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0000.png", "metadata/rollout_202/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0001.png", "metadata/rollout_202/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0002.png", "metadata/rollout_202/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0003.png", "metadata/rollout_202/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_202/agentview/0004.png", "metadata/rollout_202/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0005.png", "metadata/rollout_202/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0006.png", "metadata/rollout_202/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0007.png", "metadata/rollout_202/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0008.png", "metadata/rollout_202/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0009.png", "metadata/rollout_202/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0010.png", "metadata/rollout_202/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0011.png", "metadata/rollout_202/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0012.png", "metadata/rollout_202/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0013.png", "metadata/rollout_202/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0014.png", "metadata/rollout_202/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0015.png", "metadata/rollout_202/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0016.png", "metadata/rollout_202/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0017.png", "metadata/rollout_202/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0018.png", "metadata/rollout_202/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0019.png", "metadata/rollout_202/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0020.png", "metadata/rollout_202/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0021.png", "metadata/rollout_202/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0022.png", "metadata/rollout_202/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_202/agentview/0023.png", "metadata/rollout_202/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_202/agentview/0024.png", "metadata/rollout_202/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_202/agentview/0025.png", "metadata/rollout_202/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_202/agentview/0026.png", "metadata/rollout_202/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_202/agentview/0027.png", "metadata/rollout_202/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_202/agentview/0028.png", "metadata/rollout_202/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_202/agentview/0029.png", "metadata/rollout_202/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_202/agentview/0030.png", "metadata/rollout_202/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_202/agentview/0031.png", "metadata/rollout_202/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0032.png", "metadata/rollout_202/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_202/agentview/0033.png", "metadata/rollout_202/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_202/agentview/0034.png", "metadata/rollout_202/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_202/agentview/0035.png", "metadata/rollout_202/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_202/agentview/0036.png", "metadata/rollout_202/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_202/agentview/0037.png", "metadata/rollout_202/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0038.png", "metadata/rollout_202/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0039.png", "metadata/rollout_202/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0040.png", "metadata/rollout_202/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_202/agentview/0041.png", "metadata/rollout_202/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_202/agentview/0042.png", "metadata/rollout_202/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_202/agentview/0043.png", "metadata/rollout_202/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_202/agentview/0044.png", "metadata/rollout_202/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_202/agentview/0045.png", "metadata/rollout_202/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_BACK, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_202/agentview/0046.png", "metadata/rollout_202/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_BACK, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_202/agentview/0047.png", "metadata/rollout_202/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_202/agentview/0048.png", "metadata/rollout_202/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_202/agentview/0049.png", "metadata/rollout_202/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_202/agentview/0050.png", "metadata/rollout_202/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_202/agentview/0050.png", "metadata/rollout_202/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0000.png", "metadata/rollout_203/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0001.png", "metadata/rollout_203/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0002.png", "metadata/rollout_203/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0003.png", "metadata/rollout_203/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0004.png", "metadata/rollout_203/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0005.png", "metadata/rollout_203/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0006.png", "metadata/rollout_203/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0007.png", "metadata/rollout_203/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0008.png", "metadata/rollout_203/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0009.png", "metadata/rollout_203/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0010.png", "metadata/rollout_203/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0011.png", "metadata/rollout_203/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0012.png", "metadata/rollout_203/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0013.png", "metadata/rollout_203/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0014.png", "metadata/rollout_203/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0015.png", "metadata/rollout_203/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0016.png", "metadata/rollout_203/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0017.png", "metadata/rollout_203/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0018.png", "metadata/rollout_203/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0019.png", "metadata/rollout_203/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0020.png", "metadata/rollout_203/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0021.png", "metadata/rollout_203/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0022.png", "metadata/rollout_203/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0023.png", "metadata/rollout_203/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0024.png", "metadata/rollout_203/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0025.png", "metadata/rollout_203/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_203/agentview/0026.png", "metadata/rollout_203/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_203/agentview/0027.png", "metadata/rollout_203/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0028.png", "metadata/rollout_203/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0029.png", "metadata/rollout_203/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0030.png", "metadata/rollout_203/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0031.png", "metadata/rollout_203/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0032.png", "metadata/rollout_203/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0033.png", "metadata/rollout_203/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0034.png", "metadata/rollout_203/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0035.png", "metadata/rollout_203/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0036.png", "metadata/rollout_203/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0037.png", "metadata/rollout_203/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0038.png", "metadata/rollout_203/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0039.png", "metadata/rollout_203/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0040.png", "metadata/rollout_203/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0041.png", "metadata/rollout_203/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0042.png", "metadata/rollout_203/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0043.png", "metadata/rollout_203/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0044.png", "metadata/rollout_203/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0045.png", "metadata/rollout_203/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0046.png", "metadata/rollout_203/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0047.png", "metadata/rollout_203/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_203/agentview/0048.png", "metadata/rollout_203/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0049.png", "metadata/rollout_203/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0050.png", "metadata/rollout_203/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0051.png", "metadata/rollout_203/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0052.png", "metadata/rollout_203/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0053.png", "metadata/rollout_203/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_203/agentview/0054.png", "metadata/rollout_203/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0055.png", "metadata/rollout_203/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0056.png", "metadata/rollout_203/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0057.png", "metadata/rollout_203/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_203/agentview/0058.png", "metadata/rollout_203/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_203/agentview/0059.png", "metadata/rollout_203/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0060.png", "metadata/rollout_203/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0061.png", "metadata/rollout_203/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0062.png", "metadata/rollout_203/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_203/agentview/0063.png", "metadata/rollout_203/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_203/agentview/0063.png", "metadata/rollout_203/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0000.png", "metadata/rollout_204/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0001.png", "metadata/rollout_204/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0002.png", "metadata/rollout_204/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0003.png", "metadata/rollout_204/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0004.png", "metadata/rollout_204/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0005.png", "metadata/rollout_204/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0006.png", "metadata/rollout_204/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0007.png", "metadata/rollout_204/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_204/agentview/0008.png", "metadata/rollout_204/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_204/agentview/0009.png", "metadata/rollout_204/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_204/agentview/0010.png", "metadata/rollout_204/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0011.png", "metadata/rollout_204/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0012.png", "metadata/rollout_204/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0013.png", "metadata/rollout_204/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0014.png", "metadata/rollout_204/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0015.png", "metadata/rollout_204/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0016.png", "metadata/rollout_204/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0017.png", "metadata/rollout_204/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0018.png", "metadata/rollout_204/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0019.png", "metadata/rollout_204/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_204/agentview/0020.png", "metadata/rollout_204/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0021.png", "metadata/rollout_204/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0022.png", "metadata/rollout_204/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0023.png", "metadata/rollout_204/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_204/agentview/0024.png", "metadata/rollout_204/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_204/agentview/0025.png", "metadata/rollout_204/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_204/agentview/0026.png", "metadata/rollout_204/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_204/agentview/0027.png", "metadata/rollout_204/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_204/agentview/0028.png", "metadata/rollout_204/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_204/agentview/0029.png", "metadata/rollout_204/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_204/agentview/0030.png", "metadata/rollout_204/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0031.png", "metadata/rollout_204/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_204/agentview/0032.png", "metadata/rollout_204/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_204/agentview/0033.png", "metadata/rollout_204/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0034.png", "metadata/rollout_204/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0035.png", "metadata/rollout_204/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0036.png", "metadata/rollout_204/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0037.png", "metadata/rollout_204/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0038.png", "metadata/rollout_204/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0039.png", "metadata/rollout_204/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0040.png", "metadata/rollout_204/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0041.png", "metadata/rollout_204/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0042.png", "metadata/rollout_204/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_204/agentview/0043.png", "metadata/rollout_204/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_204/agentview/0044.png", "metadata/rollout_204/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0045.png", "metadata/rollout_204/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0046.png", "metadata/rollout_204/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0047.png", "metadata/rollout_204/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_204/agentview/0048.png", "metadata/rollout_204/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_204/agentview/0049.png", "metadata/rollout_204/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_204/agentview/0050.png", "metadata/rollout_204/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_204/agentview/0051.png", "metadata/rollout_204/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_204/agentview/0052.png", "metadata/rollout_204/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_204/agentview/0053.png", "metadata/rollout_204/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_204/agentview/0053.png", "metadata/rollout_204/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0000.png", "metadata/rollout_205/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0001.png", "metadata/rollout_205/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0002.png", "metadata/rollout_205/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0003.png", "metadata/rollout_205/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0004.png", "metadata/rollout_205/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0005.png", "metadata/rollout_205/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0006.png", "metadata/rollout_205/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0007.png", "metadata/rollout_205/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0008.png", "metadata/rollout_205/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0009.png", "metadata/rollout_205/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0010.png", "metadata/rollout_205/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0011.png", "metadata/rollout_205/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0012.png", "metadata/rollout_205/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0013.png", "metadata/rollout_205/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0014.png", "metadata/rollout_205/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0015.png", "metadata/rollout_205/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0016.png", "metadata/rollout_205/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0017.png", "metadata/rollout_205/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0018.png", "metadata/rollout_205/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0019.png", "metadata/rollout_205/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0020.png", "metadata/rollout_205/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0021.png", "metadata/rollout_205/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0022.png", "metadata/rollout_205/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0023.png", "metadata/rollout_205/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_205/agentview/0024.png", "metadata/rollout_205/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_205/agentview/0025.png", "metadata/rollout_205/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0026.png", "metadata/rollout_205/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0027.png", "metadata/rollout_205/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0028.png", "metadata/rollout_205/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0029.png", "metadata/rollout_205/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0030.png", "metadata/rollout_205/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0031.png", "metadata/rollout_205/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0032.png", "metadata/rollout_205/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0033.png", "metadata/rollout_205/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_205/agentview/0034.png", "metadata/rollout_205/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_205/agentview/0035.png", "metadata/rollout_205/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0036.png", "metadata/rollout_205/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0037.png", "metadata/rollout_205/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0038.png", "metadata/rollout_205/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_205/agentview/0039.png", "metadata/rollout_205/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_205/agentview/0040.png", "metadata/rollout_205/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0041.png", "metadata/rollout_205/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0042.png", "metadata/rollout_205/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0043.png", "metadata/rollout_205/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0044.png", "metadata/rollout_205/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0045.png", "metadata/rollout_205/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0046.png", "metadata/rollout_205/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0047.png", "metadata/rollout_205/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0048.png", "metadata/rollout_205/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_205/agentview/0049.png", "metadata/rollout_205/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_205/agentview/0050.png", "metadata/rollout_205/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_205/agentview/0051.png", "metadata/rollout_205/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0052.png", "metadata/rollout_205/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0053.png", "metadata/rollout_205/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0054.png", "metadata/rollout_205/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0055.png", "metadata/rollout_205/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_205/agentview/0056.png", "metadata/rollout_205/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_205/agentview/0057.png", "metadata/rollout_205/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_205/agentview/0057.png", "metadata/rollout_205/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0000.png", "metadata/rollout_206/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0001.png", "metadata/rollout_206/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0002.png", "metadata/rollout_206/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0003.png", "metadata/rollout_206/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0004.png", "metadata/rollout_206/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0005.png", "metadata/rollout_206/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0006.png", "metadata/rollout_206/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0007.png", "metadata/rollout_206/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0008.png", "metadata/rollout_206/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_206/agentview/0009.png", "metadata/rollout_206/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_206/agentview/0010.png", "metadata/rollout_206/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_206/agentview/0011.png", "metadata/rollout_206/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_206/agentview/0012.png", "metadata/rollout_206/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_206/agentview/0013.png", "metadata/rollout_206/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_206/agentview/0014.png", "metadata/rollout_206/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_206/agentview/0015.png", "metadata/rollout_206/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_206/agentview/0016.png", "metadata/rollout_206/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0017.png", "metadata/rollout_206/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0018.png", "metadata/rollout_206/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0019.png", "metadata/rollout_206/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0020.png", "metadata/rollout_206/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0021.png", "metadata/rollout_206/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0022.png", "metadata/rollout_206/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0023.png", "metadata/rollout_206/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0024.png", "metadata/rollout_206/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0025.png", "metadata/rollout_206/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_206/agentview/0026.png", "metadata/rollout_206/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0027.png", "metadata/rollout_206/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_206/agentview/0028.png", "metadata/rollout_206/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0029.png", "metadata/rollout_206/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0030.png", "metadata/rollout_206/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0031.png", "metadata/rollout_206/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_206/agentview/0032.png", "metadata/rollout_206/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_206/agentview/0033.png", "metadata/rollout_206/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_206/agentview/0034.png", "metadata/rollout_206/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_206/agentview/0035.png", "metadata/rollout_206/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_206/agentview/0036.png", "metadata/rollout_206/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_206/agentview/0037.png", "metadata/rollout_206/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_206/agentview/0038.png", "metadata/rollout_206/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_206/agentview/0039.png", "metadata/rollout_206/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0040.png", "metadata/rollout_206/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0041.png", "metadata/rollout_206/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0042.png", "metadata/rollout_206/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0043.png", "metadata/rollout_206/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0044.png", "metadata/rollout_206/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0045.png", "metadata/rollout_206/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0046.png", "metadata/rollout_206/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_206/agentview/0047.png", "metadata/rollout_206/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_206/agentview/0048.png", "metadata/rollout_206/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_206/agentview/0049.png", "metadata/rollout_206/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_206/agentview/0050.png", "metadata/rollout_206/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0051.png", "metadata/rollout_206/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0052.png", "metadata/rollout_206/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0053.png", "metadata/rollout_206/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0054.png", "metadata/rollout_206/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0055.png", "metadata/rollout_206/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0056.png", "metadata/rollout_206/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0057.png", "metadata/rollout_206/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0058.png", "metadata/rollout_206/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0059.png", "metadata/rollout_206/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_206/agentview/0060.png", "metadata/rollout_206/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0061.png", "metadata/rollout_206/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0062.png", "metadata/rollout_206/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0063.png", "metadata/rollout_206/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_206/agentview/0064.png", "metadata/rollout_206/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_206/agentview/0065.png", "metadata/rollout_206/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_206/agentview/0066.png", "metadata/rollout_206/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_206/agentview/0067.png", "metadata/rollout_206/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_206/agentview/0068.png", "metadata/rollout_206/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_206/agentview/0069.png", "metadata/rollout_206/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_206/agentview/0069.png", "metadata/rollout_206/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_207/agentview/0000.png", "metadata/rollout_207/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_207/agentview/0001.png", "metadata/rollout_207/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_207/agentview/0002.png", "metadata/rollout_207/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_207/agentview/0003.png", "metadata/rollout_207/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_207/agentview/0004.png", "metadata/rollout_207/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_207/agentview/0005.png", "metadata/rollout_207/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_207/agentview/0006.png", "metadata/rollout_207/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0007.png", "metadata/rollout_207/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0008.png", "metadata/rollout_207/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0009.png", "metadata/rollout_207/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0010.png", "metadata/rollout_207/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0011.png", "metadata/rollout_207/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0012.png", "metadata/rollout_207/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0013.png", "metadata/rollout_207/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0014.png", "metadata/rollout_207/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0015.png", "metadata/rollout_207/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0016.png", "metadata/rollout_207/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0017.png", "metadata/rollout_207/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0018.png", "metadata/rollout_207/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0019.png", "metadata/rollout_207/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0020.png", "metadata/rollout_207/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0021.png", "metadata/rollout_207/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0022.png", "metadata/rollout_207/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_207/agentview/0023.png", "metadata/rollout_207/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0024.png", "metadata/rollout_207/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0025.png", "metadata/rollout_207/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0026.png", "metadata/rollout_207/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0027.png", "metadata/rollout_207/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0028.png", "metadata/rollout_207/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0029.png", "metadata/rollout_207/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_207/agentview/0030.png", "metadata/rollout_207/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_207/agentview/0031.png", "metadata/rollout_207/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0032.png", "metadata/rollout_207/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0033.png", "metadata/rollout_207/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0034.png", "metadata/rollout_207/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0035.png", "metadata/rollout_207/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0036.png", "metadata/rollout_207/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0037.png", "metadata/rollout_207/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0038.png", "metadata/rollout_207/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0039.png", "metadata/rollout_207/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_207/agentview/0040.png", "metadata/rollout_207/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0041.png", "metadata/rollout_207/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0042.png", "metadata/rollout_207/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0043.png", "metadata/rollout_207/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0044.png", "metadata/rollout_207/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0045.png", "metadata/rollout_207/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0046.png", "metadata/rollout_207/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0047.png", "metadata/rollout_207/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0048.png", "metadata/rollout_207/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0049.png", "metadata/rollout_207/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_207/agentview/0050.png", "metadata/rollout_207/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0051.png", "metadata/rollout_207/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0052.png", "metadata/rollout_207/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0053.png", "metadata/rollout_207/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_207/agentview/0054.png", "metadata/rollout_207/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_207/agentview/0055.png", "metadata/rollout_207/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0056.png", "metadata/rollout_207/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0057.png", "metadata/rollout_207/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0058.png", "metadata/rollout_207/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_207/agentview/0059.png", "metadata/rollout_207/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_207/agentview/0059.png", "metadata/rollout_207/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0000.png", "metadata/rollout_208/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0001.png", "metadata/rollout_208/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0002.png", "metadata/rollout_208/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0003.png", "metadata/rollout_208/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0004.png", "metadata/rollout_208/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0005.png", "metadata/rollout_208/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0006.png", "metadata/rollout_208/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0007.png", "metadata/rollout_208/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0008.png", "metadata/rollout_208/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0009.png", "metadata/rollout_208/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0010.png", "metadata/rollout_208/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0011.png", "metadata/rollout_208/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0012.png", "metadata/rollout_208/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0013.png", "metadata/rollout_208/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0014.png", "metadata/rollout_208/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0015.png", "metadata/rollout_208/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_208/agentview/0016.png", "metadata/rollout_208/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_208/agentview/0017.png", "metadata/rollout_208/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_208/agentview/0018.png", "metadata/rollout_208/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_208/agentview/0019.png", "metadata/rollout_208/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_208/agentview/0020.png", "metadata/rollout_208/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_208/agentview/0021.png", "metadata/rollout_208/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_208/agentview/0022.png", "metadata/rollout_208/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0023.png", "metadata/rollout_208/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0024.png", "metadata/rollout_208/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0025.png", "metadata/rollout_208/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0026.png", "metadata/rollout_208/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_208/agentview/0027.png", "metadata/rollout_208/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_208/agentview/0028.png", "metadata/rollout_208/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0029.png", "metadata/rollout_208/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0030.png", "metadata/rollout_208/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0031.png", "metadata/rollout_208/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0032.png", "metadata/rollout_208/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_208/agentview/0033.png", "metadata/rollout_208/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0034.png", "metadata/rollout_208/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0035.png", "metadata/rollout_208/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0036.png", "metadata/rollout_208/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0037.png", "metadata/rollout_208/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0038.png", "metadata/rollout_208/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0039.png", "metadata/rollout_208/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_208/agentview/0040.png", "metadata/rollout_208/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_208/agentview/0041.png", "metadata/rollout_208/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_208/agentview/0042.png", "metadata/rollout_208/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0043.png", "metadata/rollout_208/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_208/agentview/0044.png", "metadata/rollout_208/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_208/agentview/0045.png", "metadata/rollout_208/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_208/agentview/0046.png", "metadata/rollout_208/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_208/agentview/0047.png", "metadata/rollout_208/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_208/agentview/0048.png", "metadata/rollout_208/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_208/agentview/0049.png", "metadata/rollout_208/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_208/agentview/0050.png", "metadata/rollout_208/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0051.png", "metadata/rollout_208/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0052.png", "metadata/rollout_208/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0053.png", "metadata/rollout_208/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0054.png", "metadata/rollout_208/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_208/agentview/0055.png", "metadata/rollout_208/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_208/agentview/0056.png", "metadata/rollout_208/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_208/agentview/0057.png", "metadata/rollout_208/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0058.png", "metadata/rollout_208/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0059.png", "metadata/rollout_208/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0060.png", "metadata/rollout_208/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0061.png", "metadata/rollout_208/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_208/agentview/0062.png", "metadata/rollout_208/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_208/agentview/0062.png", "metadata/rollout_208/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0000.png", "metadata/rollout_209/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0001.png", "metadata/rollout_209/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0002.png", "metadata/rollout_209/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0003.png", "metadata/rollout_209/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0004.png", "metadata/rollout_209/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0005.png", "metadata/rollout_209/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0006.png", "metadata/rollout_209/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0007.png", "metadata/rollout_209/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0008.png", "metadata/rollout_209/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0009.png", "metadata/rollout_209/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0010.png", "metadata/rollout_209/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0011.png", "metadata/rollout_209/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0012.png", "metadata/rollout_209/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0013.png", "metadata/rollout_209/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0014.png", "metadata/rollout_209/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0015.png", "metadata/rollout_209/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0016.png", "metadata/rollout_209/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0017.png", "metadata/rollout_209/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0018.png", "metadata/rollout_209/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0019.png", "metadata/rollout_209/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0020.png", "metadata/rollout_209/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0021.png", "metadata/rollout_209/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0022.png", "metadata/rollout_209/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0023.png", "metadata/rollout_209/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0024.png", "metadata/rollout_209/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0025.png", "metadata/rollout_209/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0026.png", "metadata/rollout_209/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0027.png", "metadata/rollout_209/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0028.png", "metadata/rollout_209/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0029.png", "metadata/rollout_209/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0030.png", "metadata/rollout_209/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0031.png", "metadata/rollout_209/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_209/agentview/0032.png", "metadata/rollout_209/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0033.png", "metadata/rollout_209/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0034.png", "metadata/rollout_209/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_209/agentview/0035.png", "metadata/rollout_209/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0036.png", "metadata/rollout_209/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0037.png", "metadata/rollout_209/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0038.png", "metadata/rollout_209/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_209/agentview/0039.png", "metadata/rollout_209/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0040.png", "metadata/rollout_209/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0041.png", "metadata/rollout_209/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0042.png", "metadata/rollout_209/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0043.png", "metadata/rollout_209/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0044.png", "metadata/rollout_209/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0045.png", "metadata/rollout_209/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0046.png", "metadata/rollout_209/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0047.png", "metadata/rollout_209/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_209/agentview/0048.png", "metadata/rollout_209/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_209/agentview/0049.png", "metadata/rollout_209/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0050.png", "metadata/rollout_209/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0051.png", "metadata/rollout_209/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0052.png", "metadata/rollout_209/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0053.png", "metadata/rollout_209/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0054.png", "metadata/rollout_209/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0055.png", "metadata/rollout_209/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0056.png", "metadata/rollout_209/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_209/agentview/0057.png", "metadata/rollout_209/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_209/agentview/0058.png", "metadata/rollout_209/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_209/agentview/0059.png", "metadata/rollout_209/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_209/agentview/0060.png", "metadata/rollout_209/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_209/agentview/0061.png", "metadata/rollout_209/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_209/agentview/0062.png", "metadata/rollout_209/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0063.png", "metadata/rollout_209/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0064.png", "metadata/rollout_209/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_209/agentview/0065.png", "metadata/rollout_209/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0066.png", "metadata/rollout_209/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0067.png", "metadata/rollout_209/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0068.png", "metadata/rollout_209/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0069.png", "metadata/rollout_209/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_209/agentview/0070.png", "metadata/rollout_209/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_209/agentview/0071.png", "metadata/rollout_209/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0072.png", "metadata/rollout_209/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0073.png", "metadata/rollout_209/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0074.png", "metadata/rollout_209/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0075.png", "metadata/rollout_209/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_209/agentview/0076.png", "metadata/rollout_209/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_209/agentview/0076.png", "metadata/rollout_209/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0000.png", "metadata/rollout_210/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0001.png", "metadata/rollout_210/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_210/agentview/0002.png", "metadata/rollout_210/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0003.png", "metadata/rollout_210/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0004.png", "metadata/rollout_210/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0005.png", "metadata/rollout_210/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0006.png", "metadata/rollout_210/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0007.png", "metadata/rollout_210/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0008.png", "metadata/rollout_210/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0009.png", "metadata/rollout_210/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0010.png", "metadata/rollout_210/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0011.png", "metadata/rollout_210/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0012.png", "metadata/rollout_210/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0013.png", "metadata/rollout_210/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0014.png", "metadata/rollout_210/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0015.png", "metadata/rollout_210/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0016.png", "metadata/rollout_210/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0017.png", "metadata/rollout_210/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0018.png", "metadata/rollout_210/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0019.png", "metadata/rollout_210/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0020.png", "metadata/rollout_210/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0021.png", "metadata/rollout_210/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0022.png", "metadata/rollout_210/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0023.png", "metadata/rollout_210/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0024.png", "metadata/rollout_210/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0025.png", "metadata/rollout_210/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_210/agentview/0026.png", "metadata/rollout_210/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0027.png", "metadata/rollout_210/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0028.png", "metadata/rollout_210/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0029.png", "metadata/rollout_210/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0030.png", "metadata/rollout_210/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0031.png", "metadata/rollout_210/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0032.png", "metadata/rollout_210/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_210/agentview/0033.png", "metadata/rollout_210/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_210/agentview/0034.png", "metadata/rollout_210/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_210/agentview/0035.png", "metadata/rollout_210/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0036.png", "metadata/rollout_210/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_210/agentview/0037.png", "metadata/rollout_210/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_210/agentview/0038.png", "metadata/rollout_210/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_210/agentview/0039.png", "metadata/rollout_210/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_210/agentview/0040.png", "metadata/rollout_210/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_210/agentview/0041.png", "metadata/rollout_210/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_210/agentview/0042.png", "metadata/rollout_210/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_210/agentview/0043.png", "metadata/rollout_210/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_210/agentview/0044.png", "metadata/rollout_210/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_210/agentview/0045.png", "metadata/rollout_210/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0046.png", "metadata/rollout_210/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0047.png", "metadata/rollout_210/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0048.png", "metadata/rollout_210/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_210/agentview/0049.png", "metadata/rollout_210/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_210/agentview/0050.png", "metadata/rollout_210/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_210/agentview/0051.png", "metadata/rollout_210/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0052.png", "metadata/rollout_210/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0053.png", "metadata/rollout_210/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0054.png", "metadata/rollout_210/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_210/agentview/0055.png", "metadata/rollout_210/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_210/agentview/0056.png", "metadata/rollout_210/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_210/agentview/0056.png", "metadata/rollout_210/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0000.png", "metadata/rollout_211/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0001.png", "metadata/rollout_211/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0002.png", "metadata/rollout_211/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0003.png", "metadata/rollout_211/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0004.png", "metadata/rollout_211/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0005.png", "metadata/rollout_211/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0006.png", "metadata/rollout_211/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0007.png", "metadata/rollout_211/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0008.png", "metadata/rollout_211/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0009.png", "metadata/rollout_211/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0010.png", "metadata/rollout_211/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0011.png", "metadata/rollout_211/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0012.png", "metadata/rollout_211/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0013.png", "metadata/rollout_211/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0014.png", "metadata/rollout_211/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0015.png", "metadata/rollout_211/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0016.png", "metadata/rollout_211/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0017.png", "metadata/rollout_211/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0018.png", "metadata/rollout_211/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0019.png", "metadata/rollout_211/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0020.png", "metadata/rollout_211/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0021.png", "metadata/rollout_211/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0022.png", "metadata/rollout_211/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0023.png", "metadata/rollout_211/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0024.png", "metadata/rollout_211/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0025.png", "metadata/rollout_211/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_211/agentview/0026.png", "metadata/rollout_211/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0027.png", "metadata/rollout_211/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_211/agentview/0028.png", "metadata/rollout_211/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0029.png", "metadata/rollout_211/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0030.png", "metadata/rollout_211/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0031.png", "metadata/rollout_211/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0032.png", "metadata/rollout_211/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_211/agentview/0033.png", "metadata/rollout_211/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0034.png", "metadata/rollout_211/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0035.png", "metadata/rollout_211/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0036.png", "metadata/rollout_211/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0037.png", "metadata/rollout_211/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0038.png", "metadata/rollout_211/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0039.png", "metadata/rollout_211/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0040.png", "metadata/rollout_211/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0041.png", "metadata/rollout_211/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_211/agentview/0042.png", "metadata/rollout_211/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_211/agentview/0043.png", "metadata/rollout_211/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0044.png", "metadata/rollout_211/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0045.png", "metadata/rollout_211/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0046.png", "metadata/rollout_211/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0047.png", "metadata/rollout_211/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0048.png", "metadata/rollout_211/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0049.png", "metadata/rollout_211/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0050.png", "metadata/rollout_211/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_211/agentview/0051.png", "metadata/rollout_211/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_211/agentview/0052.png", "metadata/rollout_211/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_211/agentview/0053.png", "metadata/rollout_211/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_211/agentview/0054.png", "metadata/rollout_211/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_211/agentview/0055.png", "metadata/rollout_211/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_211/agentview/0056.png", "metadata/rollout_211/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0057.png", "metadata/rollout_211/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0058.png", "metadata/rollout_211/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0059.png", "metadata/rollout_211/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0060.png", "metadata/rollout_211/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0061.png", "metadata/rollout_211/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0062.png", "metadata/rollout_211/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0063.png", "metadata/rollout_211/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0064.png", "metadata/rollout_211/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0065.png", "metadata/rollout_211/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0066.png", "metadata/rollout_211/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0067.png", "metadata/rollout_211/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0068.png", "metadata/rollout_211/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0069.png", "metadata/rollout_211/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0070.png", "metadata/rollout_211/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_211/agentview/0071.png", "metadata/rollout_211/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_211/agentview/0072.png", "metadata/rollout_211/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_211/agentview/0073.png", "metadata/rollout_211/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0074.png", "metadata/rollout_211/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0075.png", "metadata/rollout_211/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0076.png", "metadata/rollout_211/wrist/0076.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_211/agentview/0077.png", "metadata/rollout_211/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_211/agentview/0077.png", "metadata/rollout_211/wrist/0077.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0000.png", "metadata/rollout_212/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0001.png", "metadata/rollout_212/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0002.png", "metadata/rollout_212/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0003.png", "metadata/rollout_212/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_212/agentview/0004.png", "metadata/rollout_212/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_212/agentview/0005.png", "metadata/rollout_212/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0006.png", "metadata/rollout_212/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0007.png", "metadata/rollout_212/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0008.png", "metadata/rollout_212/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0009.png", "metadata/rollout_212/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0010.png", "metadata/rollout_212/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0011.png", "metadata/rollout_212/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0012.png", "metadata/rollout_212/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0013.png", "metadata/rollout_212/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0014.png", "metadata/rollout_212/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0015.png", "metadata/rollout_212/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0016.png", "metadata/rollout_212/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0017.png", "metadata/rollout_212/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0018.png", "metadata/rollout_212/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0019.png", "metadata/rollout_212/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_212/agentview/0020.png", "metadata/rollout_212/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0021.png", "metadata/rollout_212/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0022.png", "metadata/rollout_212/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0023.png", "metadata/rollout_212/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0024.png", "metadata/rollout_212/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_212/agentview/0025.png", "metadata/rollout_212/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0026.png", "metadata/rollout_212/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0027.png", "metadata/rollout_212/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0028.png", "metadata/rollout_212/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0029.png", "metadata/rollout_212/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0030.png", "metadata/rollout_212/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0031.png", "metadata/rollout_212/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_212/agentview/0032.png", "metadata/rollout_212/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_212/agentview/0033.png", "metadata/rollout_212/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_212/agentview/0034.png", "metadata/rollout_212/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0035.png", "metadata/rollout_212/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_212/agentview/0036.png", "metadata/rollout_212/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_212/agentview/0037.png", "metadata/rollout_212/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_212/agentview/0038.png", "metadata/rollout_212/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_212/agentview/0039.png", "metadata/rollout_212/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_212/agentview/0040.png", "metadata/rollout_212/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_212/agentview/0041.png", "metadata/rollout_212/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0042.png", "metadata/rollout_212/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0043.png", "metadata/rollout_212/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0044.png", "metadata/rollout_212/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_212/agentview/0045.png", "metadata/rollout_212/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_212/agentview/0046.png", "metadata/rollout_212/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0047.png", "metadata/rollout_212/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0048.png", "metadata/rollout_212/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0049.png", "metadata/rollout_212/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_212/agentview/0050.png", "metadata/rollout_212/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_212/agentview/0050.png", "metadata/rollout_212/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_213/agentview/0000.png", "metadata/rollout_213/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_213/agentview/0001.png", "metadata/rollout_213/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_213/agentview/0002.png", "metadata/rollout_213/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_213/agentview/0003.png", "metadata/rollout_213/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_213/agentview/0004.png", "metadata/rollout_213/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_213/agentview/0005.png", "metadata/rollout_213/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_213/agentview/0006.png", "metadata/rollout_213/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_213/agentview/0007.png", "metadata/rollout_213/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0008.png", "metadata/rollout_213/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0009.png", "metadata/rollout_213/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0010.png", "metadata/rollout_213/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0011.png", "metadata/rollout_213/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_213/agentview/0012.png", "metadata/rollout_213/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_213/agentview/0013.png", "metadata/rollout_213/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_213/agentview/0014.png", "metadata/rollout_213/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_213/agentview/0015.png", "metadata/rollout_213/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_213/agentview/0016.png", "metadata/rollout_213/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0017.png", "metadata/rollout_213/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0018.png", "metadata/rollout_213/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0019.png", "metadata/rollout_213/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0020.png", "metadata/rollout_213/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_213/agentview/0021.png", "metadata/rollout_213/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0022.png", "metadata/rollout_213/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0023.png", "metadata/rollout_213/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0024.png", "metadata/rollout_213/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_213/agentview/0025.png", "metadata/rollout_213/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_213/agentview/0026.png", "metadata/rollout_213/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_213/agentview/0027.png", "metadata/rollout_213/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_213/agentview/0028.png", "metadata/rollout_213/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_213/agentview/0029.png", "metadata/rollout_213/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_213/agentview/0030.png", "metadata/rollout_213/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0031.png", "metadata/rollout_213/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0032.png", "metadata/rollout_213/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0033.png", "metadata/rollout_213/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0034.png", "metadata/rollout_213/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0035.png", "metadata/rollout_213/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0036.png", "metadata/rollout_213/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0037.png", "metadata/rollout_213/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0038.png", "metadata/rollout_213/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0039.png", "metadata/rollout_213/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0040.png", "metadata/rollout_213/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0041.png", "metadata/rollout_213/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0042.png", "metadata/rollout_213/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0043.png", "metadata/rollout_213/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0044.png", "metadata/rollout_213/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0045.png", "metadata/rollout_213/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0046.png", "metadata/rollout_213/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0047.png", "metadata/rollout_213/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0048.png", "metadata/rollout_213/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0049.png", "metadata/rollout_213/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_213/agentview/0050.png", "metadata/rollout_213/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0051.png", "metadata/rollout_213/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0052.png", "metadata/rollout_213/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0053.png", "metadata/rollout_213/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_213/agentview/0054.png", "metadata/rollout_213/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_213/agentview/0055.png", "metadata/rollout_213/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_213/agentview/0056.png", "metadata/rollout_213/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_213/agentview/0057.png", "metadata/rollout_213/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_213/agentview/0058.png", "metadata/rollout_213/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_213/agentview/0059.png", "metadata/rollout_213/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_213/agentview/0060.png", "metadata/rollout_213/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_213/agentview/0060.png", "metadata/rollout_213/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0000.png", "metadata/rollout_214/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0001.png", "metadata/rollout_214/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0002.png", "metadata/rollout_214/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0003.png", "metadata/rollout_214/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0004.png", "metadata/rollout_214/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0005.png", "metadata/rollout_214/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0006.png", "metadata/rollout_214/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0007.png", "metadata/rollout_214/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0008.png", "metadata/rollout_214/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_214/agentview/0009.png", "metadata/rollout_214/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_214/agentview/0010.png", "metadata/rollout_214/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_214/agentview/0011.png", "metadata/rollout_214/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_214/agentview/0012.png", "metadata/rollout_214/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_214/agentview/0013.png", "metadata/rollout_214/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_214/agentview/0014.png", "metadata/rollout_214/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_214/agentview/0015.png", "metadata/rollout_214/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_214/agentview/0016.png", "metadata/rollout_214/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0017.png", "metadata/rollout_214/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0018.png", "metadata/rollout_214/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0019.png", "metadata/rollout_214/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0020.png", "metadata/rollout_214/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0021.png", "metadata/rollout_214/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0022.png", "metadata/rollout_214/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0023.png", "metadata/rollout_214/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0024.png", "metadata/rollout_214/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0025.png", "metadata/rollout_214/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0026.png", "metadata/rollout_214/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0027.png", "metadata/rollout_214/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0028.png", "metadata/rollout_214/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0029.png", "metadata/rollout_214/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0030.png", "metadata/rollout_214/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_214/agentview/0031.png", "metadata/rollout_214/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0032.png", "metadata/rollout_214/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0033.png", "metadata/rollout_214/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_214/agentview/0034.png", "metadata/rollout_214/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0035.png", "metadata/rollout_214/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0036.png", "metadata/rollout_214/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0037.png", "metadata/rollout_214/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_214/agentview/0038.png", "metadata/rollout_214/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0039.png", "metadata/rollout_214/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0040.png", "metadata/rollout_214/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0041.png", "metadata/rollout_214/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0042.png", "metadata/rollout_214/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0043.png", "metadata/rollout_214/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0044.png", "metadata/rollout_214/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0045.png", "metadata/rollout_214/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0046.png", "metadata/rollout_214/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_214/agentview/0047.png", "metadata/rollout_214/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_214/agentview/0048.png", "metadata/rollout_214/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0049.png", "metadata/rollout_214/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0050.png", "metadata/rollout_214/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0051.png", "metadata/rollout_214/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0052.png", "metadata/rollout_214/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_214/agentview/0053.png", "metadata/rollout_214/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_214/agentview/0054.png", "metadata/rollout_214/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_214/agentview/0055.png", "metadata/rollout_214/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0056.png", "metadata/rollout_214/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_BACK, MV_BACK, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0057.png", "metadata/rollout_214/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0058.png", "metadata/rollout_214/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0059.png", "metadata/rollout_214/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0060.png", "metadata/rollout_214/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0061.png", "metadata/rollout_214/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0062.png", "metadata/rollout_214/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0063.png", "metadata/rollout_214/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0064.png", "metadata/rollout_214/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_214/agentview/0065.png", "metadata/rollout_214/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_214/agentview/0066.png", "metadata/rollout_214/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0067.png", "metadata/rollout_214/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0068.png", "metadata/rollout_214/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0069.png", "metadata/rollout_214/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_214/agentview/0070.png", "metadata/rollout_214/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_214/agentview/0071.png", "metadata/rollout_214/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0072.png", "metadata/rollout_214/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0073.png", "metadata/rollout_214/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0074.png", "metadata/rollout_214/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_214/agentview/0075.png", "metadata/rollout_214/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_214/agentview/0075.png", "metadata/rollout_214/wrist/0075.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_215/agentview/0000.png", "metadata/rollout_215/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_215/agentview/0001.png", "metadata/rollout_215/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_215/agentview/0002.png", "metadata/rollout_215/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0003.png", "metadata/rollout_215/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0004.png", "metadata/rollout_215/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0005.png", "metadata/rollout_215/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_215/agentview/0006.png", "metadata/rollout_215/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_215/agentview/0007.png", "metadata/rollout_215/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_215/agentview/0008.png", "metadata/rollout_215/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_215/agentview/0009.png", "metadata/rollout_215/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_215/agentview/0010.png", "metadata/rollout_215/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_215/agentview/0011.png", "metadata/rollout_215/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_215/agentview/0012.png", "metadata/rollout_215/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_215/agentview/0013.png", "metadata/rollout_215/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_215/agentview/0014.png", "metadata/rollout_215/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0015.png", "metadata/rollout_215/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0016.png", "metadata/rollout_215/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0017.png", "metadata/rollout_215/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0018.png", "metadata/rollout_215/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0019.png", "metadata/rollout_215/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0020.png", "metadata/rollout_215/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0021.png", "metadata/rollout_215/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0022.png", "metadata/rollout_215/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0023.png", "metadata/rollout_215/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_215/agentview/0024.png", "metadata/rollout_215/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0025.png", "metadata/rollout_215/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0026.png", "metadata/rollout_215/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0027.png", "metadata/rollout_215/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0028.png", "metadata/rollout_215/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0029.png", "metadata/rollout_215/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0030.png", "metadata/rollout_215/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0031.png", "metadata/rollout_215/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0032.png", "metadata/rollout_215/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0033.png", "metadata/rollout_215/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0034.png", "metadata/rollout_215/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0035.png", "metadata/rollout_215/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0036.png", "metadata/rollout_215/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0037.png", "metadata/rollout_215/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0038.png", "metadata/rollout_215/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0039.png", "metadata/rollout_215/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0040.png", "metadata/rollout_215/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0041.png", "metadata/rollout_215/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_215/agentview/0042.png", "metadata/rollout_215/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0043.png", "metadata/rollout_215/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0044.png", "metadata/rollout_215/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0045.png", "metadata/rollout_215/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_215/agentview/0046.png", "metadata/rollout_215/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_215/agentview/0047.png", "metadata/rollout_215/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0048.png", "metadata/rollout_215/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0049.png", "metadata/rollout_215/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0050.png", "metadata/rollout_215/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_215/agentview/0051.png", "metadata/rollout_215/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_215/agentview/0051.png", "metadata/rollout_215/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_216/agentview/0000.png", "metadata/rollout_216/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_216/agentview/0001.png", "metadata/rollout_216/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_216/agentview/0002.png", "metadata/rollout_216/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_216/agentview/0003.png", "metadata/rollout_216/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0004.png", "metadata/rollout_216/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0005.png", "metadata/rollout_216/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0006.png", "metadata/rollout_216/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0007.png", "metadata/rollout_216/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_216/agentview/0008.png", "metadata/rollout_216/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_216/agentview/0009.png", "metadata/rollout_216/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0010.png", "metadata/rollout_216/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0011.png", "metadata/rollout_216/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0012.png", "metadata/rollout_216/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0013.png", "metadata/rollout_216/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0014.png", "metadata/rollout_216/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0015.png", "metadata/rollout_216/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0016.png", "metadata/rollout_216/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0017.png", "metadata/rollout_216/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_216/agentview/0018.png", "metadata/rollout_216/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_216/agentview/0019.png", "metadata/rollout_216/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0020.png", "metadata/rollout_216/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0021.png", "metadata/rollout_216/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0022.png", "metadata/rollout_216/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0023.png", "metadata/rollout_216/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0024.png", "metadata/rollout_216/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0025.png", "metadata/rollout_216/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0026.png", "metadata/rollout_216/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_216/agentview/0027.png", "metadata/rollout_216/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0028.png", "metadata/rollout_216/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0029.png", "metadata/rollout_216/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0030.png", "metadata/rollout_216/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_216/agentview/0031.png", "metadata/rollout_216/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0032.png", "metadata/rollout_216/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0033.png", "metadata/rollout_216/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0034.png", "metadata/rollout_216/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0035.png", "metadata/rollout_216/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0036.png", "metadata/rollout_216/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0037.png", "metadata/rollout_216/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0038.png", "metadata/rollout_216/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0039.png", "metadata/rollout_216/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0040.png", "metadata/rollout_216/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0041.png", "metadata/rollout_216/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_216/agentview/0042.png", "metadata/rollout_216/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0043.png", "metadata/rollout_216/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0044.png", "metadata/rollout_216/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0045.png", "metadata/rollout_216/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_216/agentview/0046.png", "metadata/rollout_216/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_216/agentview/0047.png", "metadata/rollout_216/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_216/agentview/0048.png", "metadata/rollout_216/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0049.png", "metadata/rollout_216/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0050.png", "metadata/rollout_216/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0051.png", "metadata/rollout_216/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_216/agentview/0052.png", "metadata/rollout_216/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_216/agentview/0052.png", "metadata/rollout_216/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0000.png", "metadata/rollout_217/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0001.png", "metadata/rollout_217/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0002.png", "metadata/rollout_217/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_217/agentview/0003.png", "metadata/rollout_217/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_217/agentview/0004.png", "metadata/rollout_217/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_217/agentview/0005.png", "metadata/rollout_217/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0006.png", "metadata/rollout_217/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0007.png", "metadata/rollout_217/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0008.png", "metadata/rollout_217/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0009.png", "metadata/rollout_217/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0010.png", "metadata/rollout_217/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_217/agentview/0011.png", "metadata/rollout_217/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_217/agentview/0012.png", "metadata/rollout_217/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0013.png", "metadata/rollout_217/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0014.png", "metadata/rollout_217/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0015.png", "metadata/rollout_217/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0016.png", "metadata/rollout_217/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0017.png", "metadata/rollout_217/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0018.png", "metadata/rollout_217/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0019.png", "metadata/rollout_217/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_217/agentview/0020.png", "metadata/rollout_217/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_217/agentview/0021.png", "metadata/rollout_217/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_217/agentview/0022.png", "metadata/rollout_217/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_217/agentview/0023.png", "metadata/rollout_217/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_217/agentview/0024.png", "metadata/rollout_217/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_217/agentview/0025.png", "metadata/rollout_217/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_217/agentview/0026.png", "metadata/rollout_217/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_217/agentview/0027.png", "metadata/rollout_217/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0028.png", "metadata/rollout_217/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0029.png", "metadata/rollout_217/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0030.png", "metadata/rollout_217/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0031.png", "metadata/rollout_217/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0032.png", "metadata/rollout_217/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0033.png", "metadata/rollout_217/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0034.png", "metadata/rollout_217/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0035.png", "metadata/rollout_217/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0036.png", "metadata/rollout_217/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0037.png", "metadata/rollout_217/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0038.png", "metadata/rollout_217/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0039.png", "metadata/rollout_217/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0040.png", "metadata/rollout_217/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0041.png", "metadata/rollout_217/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0042.png", "metadata/rollout_217/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_217/agentview/0043.png", "metadata/rollout_217/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_217/agentview/0044.png", "metadata/rollout_217/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0045.png", "metadata/rollout_217/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0046.png", "metadata/rollout_217/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0047.png", "metadata/rollout_217/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_217/agentview/0048.png", "metadata/rollout_217/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_217/agentview/0049.png", "metadata/rollout_217/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_217/agentview/0050.png", "metadata/rollout_217/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_217/agentview/0051.png", "metadata/rollout_217/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_217/agentview/0052.png", "metadata/rollout_217/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_217/agentview/0052.png", "metadata/rollout_217/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0000.png", "metadata/rollout_218/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0001.png", "metadata/rollout_218/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0002.png", "metadata/rollout_218/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0003.png", "metadata/rollout_218/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0004.png", "metadata/rollout_218/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0005.png", "metadata/rollout_218/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_218/agentview/0006.png", "metadata/rollout_218/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_218/agentview/0007.png", "metadata/rollout_218/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_218/agentview/0008.png", "metadata/rollout_218/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0009.png", "metadata/rollout_218/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0010.png", "metadata/rollout_218/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0011.png", "metadata/rollout_218/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0012.png", "metadata/rollout_218/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0013.png", "metadata/rollout_218/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0014.png", "metadata/rollout_218/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0015.png", "metadata/rollout_218/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0016.png", "metadata/rollout_218/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0017.png", "metadata/rollout_218/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0018.png", "metadata/rollout_218/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0019.png", "metadata/rollout_218/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0020.png", "metadata/rollout_218/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0021.png", "metadata/rollout_218/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0022.png", "metadata/rollout_218/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_218/agentview/0023.png", "metadata/rollout_218/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0024.png", "metadata/rollout_218/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0025.png", "metadata/rollout_218/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0026.png", "metadata/rollout_218/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0027.png", "metadata/rollout_218/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_218/agentview/0028.png", "metadata/rollout_218/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0029.png", "metadata/rollout_218/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0030.png", "metadata/rollout_218/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0031.png", "metadata/rollout_218/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0032.png", "metadata/rollout_218/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0033.png", "metadata/rollout_218/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0034.png", "metadata/rollout_218/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_218/agentview/0035.png", "metadata/rollout_218/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_218/agentview/0036.png", "metadata/rollout_218/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_218/agentview/0037.png", "metadata/rollout_218/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_218/agentview/0038.png", "metadata/rollout_218/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0039.png", "metadata/rollout_218/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_218/agentview/0040.png", "metadata/rollout_218/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_218/agentview/0041.png", "metadata/rollout_218/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_218/agentview/0042.png", "metadata/rollout_218/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_218/agentview/0043.png", "metadata/rollout_218/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_218/agentview/0044.png", "metadata/rollout_218/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_218/agentview/0045.png", "metadata/rollout_218/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_218/agentview/0046.png", "metadata/rollout_218/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0047.png", "metadata/rollout_218/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0048.png", "metadata/rollout_218/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0049.png", "metadata/rollout_218/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_218/agentview/0050.png", "metadata/rollout_218/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_218/agentview/0051.png", "metadata/rollout_218/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0052.png", "metadata/rollout_218/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0053.png", "metadata/rollout_218/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0054.png", "metadata/rollout_218/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_218/agentview/0055.png", "metadata/rollout_218/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_218/agentview/0055.png", "metadata/rollout_218/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0000.png", "metadata/rollout_219/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0001.png", "metadata/rollout_219/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0002.png", "metadata/rollout_219/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0003.png", "metadata/rollout_219/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0004.png", "metadata/rollout_219/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0005.png", "metadata/rollout_219/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0006.png", "metadata/rollout_219/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0007.png", "metadata/rollout_219/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_219/agentview/0008.png", "metadata/rollout_219/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_219/agentview/0009.png", "metadata/rollout_219/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_219/agentview/0010.png", "metadata/rollout_219/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_219/agentview/0011.png", "metadata/rollout_219/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_219/agentview/0012.png", "metadata/rollout_219/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_219/agentview/0013.png", "metadata/rollout_219/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_219/agentview/0014.png", "metadata/rollout_219/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0015.png", "metadata/rollout_219/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0016.png", "metadata/rollout_219/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0017.png", "metadata/rollout_219/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0018.png", "metadata/rollout_219/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0019.png", "metadata/rollout_219/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0020.png", "metadata/rollout_219/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0021.png", "metadata/rollout_219/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0022.png", "metadata/rollout_219/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0023.png", "metadata/rollout_219/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_219/agentview/0024.png", "metadata/rollout_219/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0025.png", "metadata/rollout_219/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0026.png", "metadata/rollout_219/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_219/agentview/0027.png", "metadata/rollout_219/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0028.png", "metadata/rollout_219/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0029.png", "metadata/rollout_219/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0030.png", "metadata/rollout_219/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_219/agentview/0031.png", "metadata/rollout_219/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0032.png", "metadata/rollout_219/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0033.png", "metadata/rollout_219/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0034.png", "metadata/rollout_219/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0035.png", "metadata/rollout_219/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0036.png", "metadata/rollout_219/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0037.png", "metadata/rollout_219/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_219/agentview/0038.png", "metadata/rollout_219/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_219/agentview/0039.png", "metadata/rollout_219/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_219/agentview/0040.png", "metadata/rollout_219/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_219/agentview/0041.png", "metadata/rollout_219/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0042.png", "metadata/rollout_219/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_219/agentview/0043.png", "metadata/rollout_219/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_219/agentview/0044.png", "metadata/rollout_219/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_219/agentview/0045.png", "metadata/rollout_219/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_219/agentview/0046.png", "metadata/rollout_219/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_219/agentview/0047.png", "metadata/rollout_219/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_219/agentview/0048.png", "metadata/rollout_219/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_219/agentview/0049.png", "metadata/rollout_219/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_219/agentview/0050.png", "metadata/rollout_219/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_219/agentview/0051.png", "metadata/rollout_219/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_219/agentview/0052.png", "metadata/rollout_219/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_219/agentview/0053.png", "metadata/rollout_219/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0054.png", "metadata/rollout_219/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0055.png", "metadata/rollout_219/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0056.png", "metadata/rollout_219/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_219/agentview/0057.png", "metadata/rollout_219/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_219/agentview/0058.png", "metadata/rollout_219/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0059.png", "metadata/rollout_219/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0060.png", "metadata/rollout_219/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0061.png", "metadata/rollout_219/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_219/agentview/0062.png", "metadata/rollout_219/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_219/agentview/0062.png", "metadata/rollout_219/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0000.png", "metadata/rollout_220/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0001.png", "metadata/rollout_220/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0002.png", "metadata/rollout_220/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_220/agentview/0003.png", "metadata/rollout_220/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0004.png", "metadata/rollout_220/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0005.png", "metadata/rollout_220/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0006.png", "metadata/rollout_220/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0007.png", "metadata/rollout_220/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0008.png", "metadata/rollout_220/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0009.png", "metadata/rollout_220/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0010.png", "metadata/rollout_220/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0011.png", "metadata/rollout_220/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0012.png", "metadata/rollout_220/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0013.png", "metadata/rollout_220/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0014.png", "metadata/rollout_220/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_220/agentview/0015.png", "metadata/rollout_220/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0016.png", "metadata/rollout_220/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0017.png", "metadata/rollout_220/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0018.png", "metadata/rollout_220/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0019.png", "metadata/rollout_220/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0020.png", "metadata/rollout_220/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0021.png", "metadata/rollout_220/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0022.png", "metadata/rollout_220/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0023.png", "metadata/rollout_220/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0024.png", "metadata/rollout_220/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0025.png", "metadata/rollout_220/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0026.png", "metadata/rollout_220/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0027.png", "metadata/rollout_220/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0028.png", "metadata/rollout_220/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0029.png", "metadata/rollout_220/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_220/agentview/0030.png", "metadata/rollout_220/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0031.png", "metadata/rollout_220/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0032.png", "metadata/rollout_220/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0033.png", "metadata/rollout_220/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_220/agentview/0034.png", "metadata/rollout_220/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_220/agentview/0035.png", "metadata/rollout_220/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0036.png", "metadata/rollout_220/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0037.png", "metadata/rollout_220/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0038.png", "metadata/rollout_220/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_220/agentview/0039.png", "metadata/rollout_220/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_220/agentview/0039.png", "metadata/rollout_220/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0000.png", "metadata/rollout_221/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0001.png", "metadata/rollout_221/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0002.png", "metadata/rollout_221/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0003.png", "metadata/rollout_221/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0004.png", "metadata/rollout_221/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0005.png", "metadata/rollout_221/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0006.png", "metadata/rollout_221/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0007.png", "metadata/rollout_221/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0008.png", "metadata/rollout_221/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0009.png", "metadata/rollout_221/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0010.png", "metadata/rollout_221/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0011.png", "metadata/rollout_221/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0012.png", "metadata/rollout_221/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0013.png", "metadata/rollout_221/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0014.png", "metadata/rollout_221/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0015.png", "metadata/rollout_221/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0016.png", "metadata/rollout_221/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0017.png", "metadata/rollout_221/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0018.png", "metadata/rollout_221/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0019.png", "metadata/rollout_221/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0020.png", "metadata/rollout_221/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0021.png", "metadata/rollout_221/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0022.png", "metadata/rollout_221/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0023.png", "metadata/rollout_221/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0024.png", "metadata/rollout_221/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0025.png", "metadata/rollout_221/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0026.png", "metadata/rollout_221/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0027.png", "metadata/rollout_221/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0028.png", "metadata/rollout_221/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_221/agentview/0029.png", "metadata/rollout_221/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_221/agentview/0030.png", "metadata/rollout_221/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0031.png", "metadata/rollout_221/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0032.png", "metadata/rollout_221/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0033.png", "metadata/rollout_221/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0034.png", "metadata/rollout_221/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0035.png", "metadata/rollout_221/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0036.png", "metadata/rollout_221/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0037.png", "metadata/rollout_221/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0038.png", "metadata/rollout_221/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0039.png", "metadata/rollout_221/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0040.png", "metadata/rollout_221/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0041.png", "metadata/rollout_221/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0042.png", "metadata/rollout_221/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0043.png", "metadata/rollout_221/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0044.png", "metadata/rollout_221/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0045.png", "metadata/rollout_221/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0046.png", "metadata/rollout_221/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0047.png", "metadata/rollout_221/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0048.png", "metadata/rollout_221/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_221/agentview/0049.png", "metadata/rollout_221/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0050.png", "metadata/rollout_221/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0051.png", "metadata/rollout_221/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0052.png", "metadata/rollout_221/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0053.png", "metadata/rollout_221/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0054.png", "metadata/rollout_221/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0055.png", "metadata/rollout_221/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0056.png", "metadata/rollout_221/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0057.png", "metadata/rollout_221/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0058.png", "metadata/rollout_221/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0059.png", "metadata/rollout_221/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0060.png", "metadata/rollout_221/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0061.png", "metadata/rollout_221/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0062.png", "metadata/rollout_221/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0063.png", "metadata/rollout_221/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0064.png", "metadata/rollout_221/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_221/agentview/0065.png", "metadata/rollout_221/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0066.png", "metadata/rollout_221/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0067.png", "metadata/rollout_221/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0068.png", "metadata/rollout_221/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_221/agentview/0069.png", "metadata/rollout_221/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_221/agentview/0070.png", "metadata/rollout_221/wrist/0070.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0071.png", "metadata/rollout_221/wrist/0071.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0072.png", "metadata/rollout_221/wrist/0072.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0073.png", "metadata/rollout_221/wrist/0073.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_221/agentview/0074.png", "metadata/rollout_221/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_221/agentview/0074.png", "metadata/rollout_221/wrist/0074.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_222/agentview/0000.png", "metadata/rollout_222/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_222/agentview/0001.png", "metadata/rollout_222/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_222/agentview/0002.png", "metadata/rollout_222/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_222/agentview/0003.png", "metadata/rollout_222/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_222/agentview/0004.png", "metadata/rollout_222/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_222/agentview/0005.png", "metadata/rollout_222/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_222/agentview/0006.png", "metadata/rollout_222/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_222/agentview/0007.png", "metadata/rollout_222/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_222/agentview/0008.png", "metadata/rollout_222/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_222/agentview/0009.png", "metadata/rollout_222/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_222/agentview/0010.png", "metadata/rollout_222/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_222/agentview/0011.png", "metadata/rollout_222/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_222/agentview/0012.png", "metadata/rollout_222/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0013.png", "metadata/rollout_222/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0014.png", "metadata/rollout_222/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0015.png", "metadata/rollout_222/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0016.png", "metadata/rollout_222/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0017.png", "metadata/rollout_222/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0018.png", "metadata/rollout_222/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0019.png", "metadata/rollout_222/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_222/agentview/0020.png", "metadata/rollout_222/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_222/agentview/0021.png", "metadata/rollout_222/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_222/agentview/0022.png", "metadata/rollout_222/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0023.png", "metadata/rollout_222/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0024.png", "metadata/rollout_222/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_222/agentview/0025.png", "metadata/rollout_222/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0026.png", "metadata/rollout_222/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0027.png", "metadata/rollout_222/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_222/agentview/0028.png", "metadata/rollout_222/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_222/agentview/0029.png", "metadata/rollout_222/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_222/agentview/0030.png", "metadata/rollout_222/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_222/agentview/0031.png", "metadata/rollout_222/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_222/agentview/0032.png", "metadata/rollout_222/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_222/agentview/0033.png", "metadata/rollout_222/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_222/agentview/0034.png", "metadata/rollout_222/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_222/agentview/0035.png", "metadata/rollout_222/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_222/agentview/0036.png", "metadata/rollout_222/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0037.png", "metadata/rollout_222/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_BACK, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0038.png", "metadata/rollout_222/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_BACK, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0039.png", "metadata/rollout_222/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0040.png", "metadata/rollout_222/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0041.png", "metadata/rollout_222/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0042.png", "metadata/rollout_222/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_222/agentview/0043.png", "metadata/rollout_222/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_222/agentview/0044.png", "metadata/rollout_222/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_222/agentview/0045.png", "metadata/rollout_222/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_222/agentview/0046.png", "metadata/rollout_222/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0047.png", "metadata/rollout_222/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0048.png", "metadata/rollout_222/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0049.png", "metadata/rollout_222/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0050.png", "metadata/rollout_222/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0051.png", "metadata/rollout_222/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0052.png", "metadata/rollout_222/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0053.png", "metadata/rollout_222/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0054.png", "metadata/rollout_222/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0055.png", "metadata/rollout_222/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_222/agentview/0056.png", "metadata/rollout_222/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_222/agentview/0057.png", "metadata/rollout_222/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_222/agentview/0058.png", "metadata/rollout_222/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_222/agentview/0059.png", "metadata/rollout_222/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_222/agentview/0060.png", "metadata/rollout_222/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_222/agentview/0061.png", "metadata/rollout_222/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_222/agentview/0061.png", "metadata/rollout_222/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0000.png", "metadata/rollout_223/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0001.png", "metadata/rollout_223/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0002.png", "metadata/rollout_223/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0003.png", "metadata/rollout_223/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0004.png", "metadata/rollout_223/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0005.png", "metadata/rollout_223/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0006.png", "metadata/rollout_223/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0007.png", "metadata/rollout_223/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0008.png", "metadata/rollout_223/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0009.png", "metadata/rollout_223/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0010.png", "metadata/rollout_223/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0011.png", "metadata/rollout_223/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0012.png", "metadata/rollout_223/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0013.png", "metadata/rollout_223/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0014.png", "metadata/rollout_223/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0015.png", "metadata/rollout_223/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0016.png", "metadata/rollout_223/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0017.png", "metadata/rollout_223/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0018.png", "metadata/rollout_223/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0019.png", "metadata/rollout_223/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0020.png", "metadata/rollout_223/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0021.png", "metadata/rollout_223/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0022.png", "metadata/rollout_223/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0023.png", "metadata/rollout_223/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0024.png", "metadata/rollout_223/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0025.png", "metadata/rollout_223/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0026.png", "metadata/rollout_223/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_223/agentview/0027.png", "metadata/rollout_223/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_223/agentview/0028.png", "metadata/rollout_223/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0029.png", "metadata/rollout_223/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0030.png", "metadata/rollout_223/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0031.png", "metadata/rollout_223/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0032.png", "metadata/rollout_223/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0033.png", "metadata/rollout_223/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0034.png", "metadata/rollout_223/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0035.png", "metadata/rollout_223/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0036.png", "metadata/rollout_223/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0037.png", "metadata/rollout_223/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0038.png", "metadata/rollout_223/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0039.png", "metadata/rollout_223/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0040.png", "metadata/rollout_223/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0041.png", "metadata/rollout_223/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0042.png", "metadata/rollout_223/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0043.png", "metadata/rollout_223/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_223/agentview/0044.png", "metadata/rollout_223/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0045.png", "metadata/rollout_223/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0046.png", "metadata/rollout_223/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0047.png", "metadata/rollout_223/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0048.png", "metadata/rollout_223/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0049.png", "metadata/rollout_223/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0050.png", "metadata/rollout_223/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0051.png", "metadata/rollout_223/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0052.png", "metadata/rollout_223/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0053.png", "metadata/rollout_223/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0054.png", "metadata/rollout_223/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0055.png", "metadata/rollout_223/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0056.png", "metadata/rollout_223/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0057.png", "metadata/rollout_223/wrist/0057.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0058.png", "metadata/rollout_223/wrist/0058.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0059.png", "metadata/rollout_223/wrist/0059.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_223/agentview/0060.png", "metadata/rollout_223/wrist/0060.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0061.png", "metadata/rollout_223/wrist/0061.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0062.png", "metadata/rollout_223/wrist/0062.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0063.png", "metadata/rollout_223/wrist/0063.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_223/agentview/0064.png", "metadata/rollout_223/wrist/0064.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_223/agentview/0065.png", "metadata/rollout_223/wrist/0065.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0066.png", "metadata/rollout_223/wrist/0066.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0067.png", "metadata/rollout_223/wrist/0067.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0068.png", "metadata/rollout_223/wrist/0068.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_223/agentview/0069.png", "metadata/rollout_223/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_223/agentview/0069.png", "metadata/rollout_223/wrist/0069.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0000.png", "metadata/rollout_224/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0001.png", "metadata/rollout_224/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0002.png", "metadata/rollout_224/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0003.png", "metadata/rollout_224/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0004.png", "metadata/rollout_224/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0005.png", "metadata/rollout_224/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0006.png", "metadata/rollout_224/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0007.png", "metadata/rollout_224/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0008.png", "metadata/rollout_224/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0009.png", "metadata/rollout_224/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_224/agentview/0010.png", "metadata/rollout_224/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_224/agentview/0011.png", "metadata/rollout_224/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_224/agentview/0012.png", "metadata/rollout_224/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_224/agentview/0013.png", "metadata/rollout_224/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0014.png", "metadata/rollout_224/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0015.png", "metadata/rollout_224/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0016.png", "metadata/rollout_224/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0017.png", "metadata/rollout_224/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_224/agentview/0018.png", "metadata/rollout_224/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0019.png", "metadata/rollout_224/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0020.png", "metadata/rollout_224/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0021.png", "metadata/rollout_224/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0022.png", "metadata/rollout_224/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_224/agentview/0023.png", "metadata/rollout_224/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0024.png", "metadata/rollout_224/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0025.png", "metadata/rollout_224/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0026.png", "metadata/rollout_224/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0027.png", "metadata/rollout_224/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0028.png", "metadata/rollout_224/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_224/agentview/0029.png", "metadata/rollout_224/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0030.png", "metadata/rollout_224/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0031.png", "metadata/rollout_224/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0032.png", "metadata/rollout_224/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0033.png", "metadata/rollout_224/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0034.png", "metadata/rollout_224/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0035.png", "metadata/rollout_224/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0036.png", "metadata/rollout_224/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0037.png", "metadata/rollout_224/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0038.png", "metadata/rollout_224/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0039.png", "metadata/rollout_224/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0040.png", "metadata/rollout_224/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0041.png", "metadata/rollout_224/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0042.png", "metadata/rollout_224/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0043.png", "metadata/rollout_224/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0044.png", "metadata/rollout_224/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0045.png", "metadata/rollout_224/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_224/agentview/0046.png", "metadata/rollout_224/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0047.png", "metadata/rollout_224/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0048.png", "metadata/rollout_224/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0049.png", "metadata/rollout_224/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_224/agentview/0050.png", "metadata/rollout_224/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_224/agentview/0051.png", "metadata/rollout_224/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0052.png", "metadata/rollout_224/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0053.png", "metadata/rollout_224/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0054.png", "metadata/rollout_224/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_224/agentview/0055.png", "metadata/rollout_224/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_224/agentview/0055.png", "metadata/rollout_224/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_225/agentview/0000.png", "metadata/rollout_225/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_225/agentview/0001.png", "metadata/rollout_225/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_225/agentview/0002.png", "metadata/rollout_225/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_225/agentview/0003.png", "metadata/rollout_225/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_225/agentview/0004.png", "metadata/rollout_225/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_225/agentview/0005.png", "metadata/rollout_225/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_225/agentview/0006.png", "metadata/rollout_225/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_225/agentview/0007.png", "metadata/rollout_225/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_225/agentview/0008.png", "metadata/rollout_225/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_225/agentview/0009.png", "metadata/rollout_225/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_225/agentview/0010.png", "metadata/rollout_225/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_225/agentview/0011.png", "metadata/rollout_225/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_225/agentview/0012.png", "metadata/rollout_225/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_225/agentview/0013.png", "metadata/rollout_225/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_225/agentview/0014.png", "metadata/rollout_225/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0015.png", "metadata/rollout_225/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0016.png", "metadata/rollout_225/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0017.png", "metadata/rollout_225/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0018.png", "metadata/rollout_225/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0019.png", "metadata/rollout_225/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0020.png", "metadata/rollout_225/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0021.png", "metadata/rollout_225/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0022.png", "metadata/rollout_225/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0023.png", "metadata/rollout_225/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0024.png", "metadata/rollout_225/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0025.png", "metadata/rollout_225/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0026.png", "metadata/rollout_225/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_225/agentview/0027.png", "metadata/rollout_225/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_225/agentview/0028.png", "metadata/rollout_225/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0029.png", "metadata/rollout_225/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0030.png", "metadata/rollout_225/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0031.png", "metadata/rollout_225/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0032.png", "metadata/rollout_225/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0033.png", "metadata/rollout_225/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0034.png", "metadata/rollout_225/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_225/agentview/0035.png", "metadata/rollout_225/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_225/agentview/0036.png", "metadata/rollout_225/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0037.png", "metadata/rollout_225/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_BACK, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_225/agentview/0038.png", "metadata/rollout_225/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_225/agentview/0039.png", "metadata/rollout_225/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_225/agentview/0040.png", "metadata/rollout_225/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_225/agentview/0041.png", "metadata/rollout_225/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_BACK, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_225/agentview/0042.png", "metadata/rollout_225/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_BACK, MV_BACK, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_225/agentview/0043.png", "metadata/rollout_225/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_BACK, MV_BACK, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_225/agentview/0044.png", "metadata/rollout_225/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_225/agentview/0045.png", "metadata/rollout_225/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_225/agentview/0046.png", "metadata/rollout_225/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0047.png", "metadata/rollout_225/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0048.png", "metadata/rollout_225/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0049.png", "metadata/rollout_225/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_225/agentview/0050.png", "metadata/rollout_225/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_BACK", "images": [ "metadata/rollout_225/agentview/0051.png", "metadata/rollout_225/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_225/agentview/0052.png", "metadata/rollout_225/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0053.png", "metadata/rollout_225/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_BACK, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0054.png", "metadata/rollout_225/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_BACK, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0055.png", "metadata/rollout_225/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_BACK, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_225/agentview/0056.png", "metadata/rollout_225/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_BACK\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_225/agentview/0056.png", "metadata/rollout_225/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0000.png", "metadata/rollout_226/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0001.png", "metadata/rollout_226/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0002.png", "metadata/rollout_226/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0003.png", "metadata/rollout_226/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0004.png", "metadata/rollout_226/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0005.png", "metadata/rollout_226/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0006.png", "metadata/rollout_226/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_226/agentview/0007.png", "metadata/rollout_226/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_226/agentview/0008.png", "metadata/rollout_226/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_226/agentview/0009.png", "metadata/rollout_226/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_226/agentview/0010.png", "metadata/rollout_226/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_226/agentview/0011.png", "metadata/rollout_226/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_226/agentview/0012.png", "metadata/rollout_226/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0013.png", "metadata/rollout_226/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0014.png", "metadata/rollout_226/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0015.png", "metadata/rollout_226/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0016.png", "metadata/rollout_226/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0017.png", "metadata/rollout_226/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0018.png", "metadata/rollout_226/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0019.png", "metadata/rollout_226/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0020.png", "metadata/rollout_226/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0021.png", "metadata/rollout_226/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0022.png", "metadata/rollout_226/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0023.png", "metadata/rollout_226/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0024.png", "metadata/rollout_226/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_226/agentview/0025.png", "metadata/rollout_226/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_226/agentview/0026.png", "metadata/rollout_226/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0027.png", "metadata/rollout_226/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0028.png", "metadata/rollout_226/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0029.png", "metadata/rollout_226/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0030.png", "metadata/rollout_226/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0031.png", "metadata/rollout_226/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0032.png", "metadata/rollout_226/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0033.png", "metadata/rollout_226/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_226/agentview/0034.png", "metadata/rollout_226/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0035.png", "metadata/rollout_226/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_226/agentview/0036.png", "metadata/rollout_226/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_226/agentview/0037.png", "metadata/rollout_226/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0038.png", "metadata/rollout_226/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_226/agentview/0039.png", "metadata/rollout_226/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_226/agentview/0040.png", "metadata/rollout_226/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_226/agentview/0041.png", "metadata/rollout_226/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_226/agentview/0042.png", "metadata/rollout_226/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_226/agentview/0043.png", "metadata/rollout_226/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_226/agentview/0044.png", "metadata/rollout_226/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_226/agentview/0045.png", "metadata/rollout_226/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0046.png", "metadata/rollout_226/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0047.png", "metadata/rollout_226/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0048.png", "metadata/rollout_226/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0049.png", "metadata/rollout_226/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_226/agentview/0050.png", "metadata/rollout_226/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_226/agentview/0051.png", "metadata/rollout_226/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0052.png", "metadata/rollout_226/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0053.png", "metadata/rollout_226/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0054.png", "metadata/rollout_226/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0055.png", "metadata/rollout_226/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_226/agentview/0056.png", "metadata/rollout_226/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_226/agentview/0056.png", "metadata/rollout_226/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_227/agentview/0000.png", "metadata/rollout_227/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_227/agentview/0001.png", "metadata/rollout_227/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_227/agentview/0002.png", "metadata/rollout_227/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_227/agentview/0003.png", "metadata/rollout_227/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0004.png", "metadata/rollout_227/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0005.png", "metadata/rollout_227/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0006.png", "metadata/rollout_227/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0007.png", "metadata/rollout_227/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_227/agentview/0008.png", "metadata/rollout_227/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_227/agentview/0009.png", "metadata/rollout_227/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_227/agentview/0010.png", "metadata/rollout_227/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_227/agentview/0011.png", "metadata/rollout_227/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0012.png", "metadata/rollout_227/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0013.png", "metadata/rollout_227/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0014.png", "metadata/rollout_227/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0015.png", "metadata/rollout_227/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0016.png", "metadata/rollout_227/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0017.png", "metadata/rollout_227/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0018.png", "metadata/rollout_227/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0019.png", "metadata/rollout_227/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_227/agentview/0020.png", "metadata/rollout_227/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0021.png", "metadata/rollout_227/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0022.png", "metadata/rollout_227/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0023.png", "metadata/rollout_227/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0024.png", "metadata/rollout_227/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0025.png", "metadata/rollout_227/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0026.png", "metadata/rollout_227/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0027.png", "metadata/rollout_227/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0028.png", "metadata/rollout_227/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0029.png", "metadata/rollout_227/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0030.png", "metadata/rollout_227/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_LEFT, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0031.png", "metadata/rollout_227/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0032.png", "metadata/rollout_227/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0033.png", "metadata/rollout_227/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0034.png", "metadata/rollout_227/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0035.png", "metadata/rollout_227/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0036.png", "metadata/rollout_227/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0037.png", "metadata/rollout_227/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0038.png", "metadata/rollout_227/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_227/agentview/0039.png", "metadata/rollout_227/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0040.png", "metadata/rollout_227/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0041.png", "metadata/rollout_227/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0042.png", "metadata/rollout_227/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_227/agentview/0043.png", "metadata/rollout_227/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_227/agentview/0044.png", "metadata/rollout_227/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0045.png", "metadata/rollout_227/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0046.png", "metadata/rollout_227/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0047.png", "metadata/rollout_227/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_227/agentview/0048.png", "metadata/rollout_227/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_227/agentview/0048.png", "metadata/rollout_227/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0000.png", "metadata/rollout_228/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0001.png", "metadata/rollout_228/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0002.png", "metadata/rollout_228/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0003.png", "metadata/rollout_228/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0004.png", "metadata/rollout_228/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0005.png", "metadata/rollout_228/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0006.png", "metadata/rollout_228/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0007.png", "metadata/rollout_228/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0008.png", "metadata/rollout_228/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0009.png", "metadata/rollout_228/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0010.png", "metadata/rollout_228/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0011.png", "metadata/rollout_228/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0012.png", "metadata/rollout_228/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0013.png", "metadata/rollout_228/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_228/agentview/0014.png", "metadata/rollout_228/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_228/agentview/0015.png", "metadata/rollout_228/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_228/agentview/0016.png", "metadata/rollout_228/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_228/agentview/0017.png", "metadata/rollout_228/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_228/agentview/0018.png", "metadata/rollout_228/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_228/agentview/0019.png", "metadata/rollout_228/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0020.png", "metadata/rollout_228/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0021.png", "metadata/rollout_228/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0022.png", "metadata/rollout_228/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0023.png", "metadata/rollout_228/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0024.png", "metadata/rollout_228/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0025.png", "metadata/rollout_228/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_228/agentview/0026.png", "metadata/rollout_228/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0027.png", "metadata/rollout_228/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_FWD, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0028.png", "metadata/rollout_228/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_FWD, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0029.png", "metadata/rollout_228/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_FWD, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0030.png", "metadata/rollout_228/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0031.png", "metadata/rollout_228/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0032.png", "metadata/rollout_228/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0033.png", "metadata/rollout_228/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0034.png", "metadata/rollout_228/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_228/agentview/0035.png", "metadata/rollout_228/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_FWD, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0036.png", "metadata/rollout_228/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0037.png", "metadata/rollout_228/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_228/agentview/0038.png", "metadata/rollout_228/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_228/agentview/0039.png", "metadata/rollout_228/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_228/agentview/0040.png", "metadata/rollout_228/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_228/agentview/0041.png", "metadata/rollout_228/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_228/agentview/0042.png", "metadata/rollout_228/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_228/agentview/0043.png", "metadata/rollout_228/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_228/agentview/0044.png", "metadata/rollout_228/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0045.png", "metadata/rollout_228/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0046.png", "metadata/rollout_228/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0047.png", "metadata/rollout_228/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_228/agentview/0048.png", "metadata/rollout_228/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_228/agentview/0049.png", "metadata/rollout_228/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0050.png", "metadata/rollout_228/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0051.png", "metadata/rollout_228/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0052.png", "metadata/rollout_228/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_228/agentview/0053.png", "metadata/rollout_228/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_228/agentview/0053.png", "metadata/rollout_228/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: none\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0000.png", "metadata/rollout_229/wrist/0000.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0001.png", "metadata/rollout_229/wrist/0001.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0002.png", "metadata/rollout_229/wrist/0002.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0003.png", "metadata/rollout_229/wrist/0003.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0004.png", "metadata/rollout_229/wrist/0004.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0005.png", "metadata/rollout_229/wrist/0005.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_229/agentview/0006.png", "metadata/rollout_229/wrist/0006.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_FWD, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_229/agentview/0007.png", "metadata/rollout_229/wrist/0007.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_229/agentview/0008.png", "metadata/rollout_229/wrist/0008.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_229/agentview/0009.png", "metadata/rollout_229/wrist/0009.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_229/agentview/0010.png", "metadata/rollout_229/wrist/0010.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_229/agentview/0011.png", "metadata/rollout_229/wrist/0011.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0012.png", "metadata/rollout_229/wrist/0012.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0013.png", "metadata/rollout_229/wrist/0013.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0014.png", "metadata/rollout_229/wrist/0014.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0015.png", "metadata/rollout_229/wrist/0015.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0016.png", "metadata/rollout_229/wrist/0016.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0017.png", "metadata/rollout_229/wrist/0017.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0018.png", "metadata/rollout_229/wrist/0018.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0019.png", "metadata/rollout_229/wrist/0019.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0020.png", "metadata/rollout_229/wrist/0020.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_229/agentview/0021.png", "metadata/rollout_229/wrist/0021.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_229/agentview/0022.png", "metadata/rollout_229/wrist/0022.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_RIGHT", "images": [ "metadata/rollout_229/agentview/0023.png", "metadata/rollout_229/wrist/0023.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0024.png", "metadata/rollout_229/wrist/0024.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0025.png", "metadata/rollout_229/wrist/0025.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0026.png", "metadata/rollout_229/wrist/0026.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0027.png", "metadata/rollout_229/wrist/0027.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "GRASP", "images": [ "metadata/rollout_229/agentview/0028.png", "metadata/rollout_229/wrist/0028.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD, MV_RIGHT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0029.png", "metadata/rollout_229/wrist/0029.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0030.png", "metadata/rollout_229/wrist/0030.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0031.png", "metadata/rollout_229/wrist/0031.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0032.png", "metadata/rollout_229/wrist/0032.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0033.png", "metadata/rollout_229/wrist/0033.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0034.png", "metadata/rollout_229/wrist/0034.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_229/agentview/0035.png", "metadata/rollout_229/wrist/0035.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_UP, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0036.png", "metadata/rollout_229/wrist/0036.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_UP, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0037.png", "metadata/rollout_229/wrist/0037.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_FWD, MV_LEFT, MV_UP, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_229/agentview/0038.png", "metadata/rollout_229/wrist/0038.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_DOWN, MV_FWD, MV_LEFT, MV_UP\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_229/agentview/0039.png", "metadata/rollout_229/wrist/0039.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_DOWN, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0040.png", "metadata/rollout_229/wrist/0040.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_FWD", "images": [ "metadata/rollout_229/agentview/0041.png", "metadata/rollout_229/wrist/0041.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_FWD, MV_FWD, MV_LEFT, MV_LEFT, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_229/agentview/0042.png", "metadata/rollout_229/wrist/0042.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_FWD, MV_FWD, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_229/agentview/0043.png", "metadata/rollout_229/wrist/0043.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_FWD, MV_FWD, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_229/agentview/0044.png", "metadata/rollout_229/wrist/0044.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_229/agentview/0045.png", "metadata/rollout_229/wrist/0045.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_FWD\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_229/agentview/0046.png", "metadata/rollout_229/wrist/0046.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_LEFT", "images": [ "metadata/rollout_229/agentview/0047.png", "metadata/rollout_229/wrist/0047.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0048.png", "metadata/rollout_229/wrist/0048.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0049.png", "metadata/rollout_229/wrist/0049.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0050.png", "metadata/rollout_229/wrist/0050.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_DOWN", "images": [ "metadata/rollout_229/agentview/0051.png", "metadata/rollout_229/wrist/0051.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "RELEASE", "images": [ "metadata/rollout_229/agentview/0052.png", "metadata/rollout_229/wrist/0052.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN, MV_LEFT\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0053.png", "metadata/rollout_229/wrist/0053.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_DOWN, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0054.png", "metadata/rollout_229/wrist/0054.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_DOWN, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0055.png", "metadata/rollout_229/wrist/0055.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_DOWN, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "MV_UP", "images": [ "metadata/rollout_229/agentview/0056.png", "metadata/rollout_229/wrist/0056.png" ] }, { "instruction": "You are controlling a robot arm with two cameras:\n- Agentview: overhead view of the robot and workspace\n- Wristview: close-up view from the gripper\n\nTask: pick up the orange block and place it on the coaster\nRecent moves, newest first: MV_UP, MV_UP, MV_UP, MV_UP, MV_DOWN\n\nOutput exactly one action token:\nMV_FWD, MV_BACK, MV_LEFT, MV_RIGHT, MV_UP, MV_DOWN, GRASP, RELEASE, DONE\n\nCheck both camera views and choose the next action:\n- Use AgentView to locate the target when it is not in wrist view\n- Use wrist view to fine-align when target is visible up close\n- GRASP when gripper fingers are aligned around the object\n- RELEASE when object is above the destination\n- DONE when the task is complete: the object is at its destination and the gripper is clear\n- Avoid repeating a direction that conflicts with the most recent move\n\nReturn the single token only, no punctuation, no explanation:", "input": "", "output": "DONE", "images": [ "metadata/rollout_229/agentview/0056.png", "metadata/rollout_229/wrist/0056.png" ] } ]